Delphic Alpha: Alpha Reports

Alpha Combinations, Part 2: From Theory to Practice

oracle — Tue, 05 May 2026 12:07:53 GMT

In Part 1 we derived the closed-form Sharpe for combining two alphas, proved the decision rule S₂ > ρ · S₁, extended it to N assets, and showed that 4–6 well-chosen signals capture 90%+ of the achievable Sharpe. All theory, one formula, zero fitted parameters.

This post is the empirical half. Using the 12 signals from the 30-Minute Intraday Alpha Atlas as a worked example, we take all 330 alpha pairs across 5 asset classes and ask: what happens when you actually fit weights to data? Four methods (equal weight, inverse variance, LASSO, Ridge), tested in- and out-of-sample, with the Part 1 maths as the benchmark.

The punchline:naive equal-weight is surprisingly hard to beat. Fitted methods improve in-sample but give some back out-of-sample. The Part 1 formula predicted this: the N-asset ceiling at ρ̄ = 0.46 leaves little room for optimisation to add value.

Which Pairs Create the Most Value?

Part 1’s decision rule says: combine when S₂ > ρ · S₁. The marginal improvement heatmap tests this empirically. For each pair, it shows how much the combination’s OOS Sharpe exceeds the better individual alpha:

Shown for Bond Futures (highest average OOS Sharpe). Diagonal: individual Sharpe. Off-diagonal: marginal improvement from optimal pairing.

The biggest improvements come from:

Cross-family pairs. Breakout + mean-reversion pairs consistently show 0.2–0.4 Sharpe improvement. These are genuinely different bets, and Part 1’s formula explains why: low ρ between families means the denominator √(1−ρ²) stays close to 1, maximising the diversification amplifier.
Long + short lookback within a family. CS-120-bar paired with CS-5-bar. Different horizons capture different dynamics while sharing signal logic. The greedy selection algorithm from Part 1 picks exactly these pairs first.
Adjacent lookbacks show near-zero improvement. F5-bar + F20-bar are too correlated (ρ > 0.8). Part 1’s destructive interference analysis applies: at equal weights, the weaker lookback dilutes the stronger one.

Top Pairs: Equity Curves

The 6 pairs with the highest OOS Sharpe, showing individual alphas (grey), optimal combination (green), and equal weight (blue dashed):

Each panel shows IS and predicted Sharpe, OOS Sharpe, and the pairwise correlation. Vertical dashed line: IS/OOS split at March 2024.

FX dominates the top 6, all involving Fade 5-bar breakout (IS Sharpe 2.4). The optimal combination barely improves on the best individual, because Part 1’s equal-Sharpe formula S·√(2/(1+ρ)) shows the multiplier is small when one alpha already dominates. The real combination gains come from weaker-but-uncorrelated pairs.

From Pairs to Full Portfolios

Pairwise analysis gives intuition, but in practice you want to combine all 12 alphas at once. Part 1 showed the theoretical ceiling: at ρ̄ = 0.46, even optimal weights can only reach ≈1.47× a single alpha’s Sharpe. How close do fitted methods get? Four approaches, ordered from simplest to most complex:

Equal Weight (1/N). Every alpha gets the same weight. Zero fitting, zero estimation error. Part 1’s formula says this achieves S·√(12/6.06) ≈ 1.41S.
Inverse-Variance. Weight inversely proportional to historical variance. Risk-parity lite: high-vol alphas get less capital.
L1 / LASSO (sparse). Maximise IS Sharpe with an L1 penalty. The L1 norm drives weak weights to exactly zero. Automatic signal selection, analogous to Part 1’s greedy algorithm but fitted to data.
L2 / Ridge (smooth). Same but with an L2 penalty. L2 never zeros a weight; it shrinks all weights proportionally.

The key distinction: L1 is a model selector (picks winners, discards the rest). L2 is a model shrinker (keeps everyone, trusts no one completely). From Part 1’s greedy analysis, we know 4–6 alphas capture most of the value, which suggests L1’s selection might be appropriate.

L1 vs L2: A Visual Intuition

Both add a penalty to the optimisation:

L1: minimise loss + λ · sum(|w_i|)
L2: minimise loss + λ · sum(w_i²)

The geometry explains everything. The L1 constraint region is a diamond; the optimal point lands on a corner where coordinates are exactly zero. The L2 constraint region is a circle, so no weight is forced to zero.

L1 says: "some of your alphas are noise. I’ll find the best 4–5 and ignore the rest."
L2 says: "all your alphas might have some signal. I’ll keep them all but shrink the uncertain ones."

Neither is universally better. L1 wins when you have many noise alphas and a few strong ones. L2 wins when many alphas carry small but real signal. From Part 1’s greedy analysis, most of our 12 alphas have positive signal, so L2 should dominate. Let’s see.

What the Weights Look Like

Weights fitted on in-sample data. Each row: one of 12 alphas. Each column: a combination method.

Equal weight is uniform: every alpha gets 1/12 = 0.083. Same weight for the best alpha (Sharpe 2.85) and the worst (−0.47).
L1 zeroes out ~half the alphas. LASSO keeps 5–6 and sets the rest to exactly zero. Compare to Part 1’s greedy selection, which also found 4–6 optimal. The methods agree on the number, but not always the selection.
L2 keeps all 12 but shrinks aggressively. No weight is zero, but weaker alphas get pushed close. Consistent with the N-asset formula: at ρ̄ = 0.46, including all 12 still adds marginal diversification.

The IS vs OOS Test

Light bars: in-sample. Solid bars: out-of-sample.

Three findings:

Equal weight is remarkably robust. Never the highest IS Sharpe, but the smallest IS-to-OOS shrinkage. Zero estimation error means nothing to overfit. Part 1 predicted this: S·√(12/6.06) requires no fitted parameters.
L1 overfits the most. Highest IS Sharpe by concentrating on IS winners, but the largest IS-to-OOS drop. The greedy algorithm from Part 1 selects on true Sharpes; L1 selects on estimated Sharpes. The estimation error in the selection step is the killer.
L2 sits in between. Shrinkage reduces overfitting vs L1 but still introduces estimation error from fitting weights. Consistent with Part 1’s observation that most alphas carry real signal.

Tuning the Penalty

IS vs OOS Sharpe as penalty strength increases. Left: LASSO with number of non-zero weights (grey). Right: Ridge.

L1: as λ increases, alphas get zeroed one by one. OOS initially improves then drops. Sweet spot: 4–6 alphas, exactly matching Part 1’s greedy selection and the N-asset ceiling.
L2: OOS Sharpe is much flatter across λ values. Ridge is less sensitive to the exact penalty choice.
The IS-OOS gap is the overfitting tax. Part 1’s formula, which uses true Sharpes and correlations, has no such gap. The gap IS the cost of estimation.

Equity Curves

Cumulative PnL. Dotted vertical line: IS/OOS boundary.

Best single alpha (grey) has the steepest slope sometimes but the deepest drawdowns.
Equal weight (blue) is smoother: shallower drawdowns, faster recovery. The diversification premium from Part 1’s √(N/(1+(N−1)ρ̄)) multiplier in action.
L1 and L2 sit between single-alpha and equal-weight. L1 partially diversifies but selection partially undoes the benefit.

Going Cross-Asset

Within-asset combination is layer one. Layer two: combine the best per-asset portfolio across all 5 asset classes.

12 alphas × 5 asset classes = 60 return streams. Cross-group correlations are structurally low. Apply Part 1’s N-asset formula at the cross-asset level: with 5 near-uncorrelated blocks, the multiplier is close to √5 ≈ 2.24×. Two-stage diversification is more robust than fitting a single 60-weight vector.

Measuring the Diversification Benefit

The diversification ratio quantifies how much portfolio vol is reduced vs the average individual alpha vol:

DR = sum(w_i * sigma_i) / sigma_portfolio

DR = 1.0 means no benefit. Higher is better. Part 1 gives the theoretical maximum: if alphas were independent, DR = √12 ≈ 3.46. At ρ̄ = 0.46, the ceiling is lower.

Even a DR of 1.5 means 33% lower portfolio vol than the average alpha. This is the empirical manifestation of Part 1’s √(2/(1+ρ)) multiplier, extended to 12 assets.

Takeaways

Start with equal weight. Zero estimation error means nothing to overfit. Part 1’s formula tells you the ceiling: S·√(N/(1+(N−1)ρ̄)). Only move to fitted weights with a specific reason and enough data.
Screen pairs with the closed-form formula. If S₂ < ρ · S₁ after a 25% haircut, the pair isn’t worth the estimation risk. Cross-family pairs beat within-family because ρ is structurally lower.
4–6 alphas capture 90%+ of the diversification benefit. Both Part 1’s greedy algorithm and L1 converge on this. Adding more of the same type hits the N-asset ceiling at ρ̄ = 0.46.
Prefer L2 over L1. Most alphas carry real signal. Ridge keeps all and shrinks; LASSO discards and overfits the selection step. Combine within-asset first, then cross-asset.

Data: this is an example implementation using the 30-Minute Intraday Alpha Atlas signals. IS/OOS split at March 2024. 330 pairwise combinations (66 per asset class × 5). L1/L2 weights fitted via sklearn Lasso/Ridge on standardised IS returns. No costs, no transaction model.

The Exotic Alpha Atlas: Beyond Textbook Signals on 30-Minute Bars

oracle — Tue, 28 Apr 2026 17:26:30 GMT

The original Alpha Atlas tested four textbook signals — breakout, MA crossover, cross-sectional momentum, and z-score reversion. Every quant has coded these at some point. But practitioners don’t stop at textbook signals. They trade distributional shape, volatility envelopes, price curvature, and return-space normalizations. Do any of these survive on intraday bars?

This post applies 18 exotic alphas from 6 non-standard signal families to the same 5 asset classes on 30-minute bars. Same rules: no optimization, no fitting, no lookahead, no transaction costs. All signals use delay_1(close) to avoid shared-variable bias.

The question: does going from textbook to exotic buy you alpha, or just complexity?

Signal Definitions

18 alphas from 6 exotic families, each at three lookback horizons N ∈ {5, 20, 120} bars (~2.5 hours, ~10 hours, ~60 hours):

Skewness Fade. Signal = −skewₙ(P). The rolling skewness of lagged close prices. Positive skewness means the right tail is fatter — prices have been making outsized moves upward. This alpha fades that asymmetry: sell when skew > 0, buy when skew < 0. At N=5 this captures micro-distribution shifts over 2.5 hours; at N=120 it captures multi-day skewness regimes.
Keltner Reversion. Signal = −(P − SMAₙ) / ATRₙ. The classic z-score mean reversion, but normalized by Average True Range instead of standard deviation. ATR uses the true range (max of |H−L|, |H−C₁|, |L−C₁|), which captures intrabar volatility and gaps that σ misses. Think of it as measuring distance from the SMA in ‘volatility envelope’ units rather than standard deviation units.
Acceleration Fade. Signal = −zscoreₙ(Δ²P). The z-scored second finite difference of lagged close: Δ²P = (Pₜ₋₁ − Pₜ₋₂) − (Pₜ₋₂ − Pₜ₋₃). This is price curvature — positive when the move is accelerating upward (convex), negative when decelerating. The alpha fades unusual acceleration: bet that parabolic moves don’t sustain. At N=5, this fires on micro-parabolas within a single session.
Return Z-Score Reversion. Signal = −zscoreₙ(rₜ₋₁) where rₜ₋₁ = Pₜ₋₁ / Pₜ₋₂. Classic MR z-scores the price level: “is the price far from its average?” This z-scores the return: “was the last return unusually large?” The two diverge in trends — price z-score stays high but return z-score resets quickly — and agree on spikes. This alpha is a pure return-space mean reversion signal.
Signed Kurtosis Fade. Signal = −skewₙ × kurtₙ. The product of negated skewness and kurtosis. When the distribution is both skewed (directional asymmetry) and fat-tailed (extreme moves likely), the signal fires hardest. Positive skew + high kurtosis = strong short. Negative skew + high kurtosis = strong long. Normal kurtosis = weak signal regardless of skew. This is a third-moment × fourth-moment interaction — no standard alpha uses this combination.
Fractal Regime Trend. Signal = (rangeₙ / pathₙ) × zscoreₙ. The ratio of the N-bar price range to the sum of absolute bar-to-bar moves is a fractal dimension proxy. High ratio ≈ price moved in a straight line (trending). Low ratio ≈ price zigzagged (choppy). Multiply by the z-score and you get a signal that trend-follows only when the market geometry says it’s actually trending. In choppy regimes the signal self-mutes. This is a regime-adaptive alpha with zero fitted parameters.

Applied identically to 30 instruments across 5 asset classes: equity index futures (ES, NQ, YM, NKD, FTSE, DAX), bond futures (TN, UB, US, ZN, ZF, ZT), commodity futures (GC, SI, CL, HG, ZW, ZC), G7 FX, and crypto (BTC, ETH, SOL, XRP, DOGE, ADA).

Period: March 2022 – February 2026. Sharpe ratios annualised from monthly returns. ~50,000 bars per instrument.

Subscribe now

The Exotic Alpha × Asset Class Heatmap

Each cell: annualised Sharpe (left) and return per trade per unit bet (right) for one exotic alpha–asset class pair:

Left: annualised Sharpe. Right: RPT — return per trade assuming a ±1 unit position. All signals use delay_1(close). No costs, no fitting, no lookahead.

What patterns emerge:

Keltner Reversion is the dominant exotic family. It shares the same directional thesis as textbook MR (fade deviations from the moving average) but normalizes by true range. If ATR normalization produces higher Sharpes than σ normalization, it tells us that intraday volatility structure (gaps, intrabar range) contains information that standard deviation ignores.
Skewness Fade separates asset classes. Markets with persistent skewness regimes (commodities trending, crypto trending) should show signal in the skewness family. Markets with symmetric microstructure (bonds, FX) may not. The heatmap reveals where distributional shape adds value beyond price levels.
Acceleration Fade captures a different effect. The second derivative is orthogonal to both level (MR) and first-derivative (momentum) signals. Parabolic move fading should work best in markets prone to overshoot-and-correct dynamics — crypto and commodities during trend exhaustion.
Return Z-Score vs Price Z-Score: the subtle difference. Return-space MR won’t flag a smooth, persistent trend (returns are normal, just consistently positive). It only fires on spikes. Compare its heatmap column with the textbook MR from the original Atlas to see where this distinction matters.

Best Exotic Alpha per Asset Class

The key comparison: how do these exotic winners stack up against the textbook winners from the original Atlas? If the best exotic alpha in an asset class beats the best textbook alpha, the added complexity pays for itself. If not, stick with the simpler signal.

Exotic vs Textbook: Does Complexity Pay?

The original 30-minute Atlas found that Fade 5-bar Breakout dominated at |Sharpe| 2.3–2.9 across bonds, FX, and commodities. That’s a ternary signal with zero parameters. The exotic families here are more sophisticated: skewness uses the third statistical moment, Keltner uses OHLC data (not just close), acceleration uses second derivatives, and return z-score uses a different normalization space.

The question isn’t whether exotic signals work — most reasonable signals have some predictive power on 50,000 bars. The question is whether they work better than the simplest version, and whether the improvement survives out of sample.

Keltner vs plain MR: ATR normalization should matter most in markets with session breaks and gaps — futures and FX rather than 24/7 crypto. The relative improvement (if any) tells us how much information lives in the intrabar range.
Acceleration vs Fade-BO: Both are mean-reverting at extremes, but Fade-BO fires at range boundaries (Donchian channel) while Acceleration fires when the move is curving. They should agree on V-shaped reversals but disagree on gradual trends.
Skewness Fade vs everything: This is the most exotic signal. If it works, it captures something genuinely new — distributional asymmetry that price levels and returns don’t measure. If it doesn’t, the third moment is noise at 30-minute frequency.

The Overfitting Tax: Exotic Edition

Same split as the original Atlas: March 2024 — 2 years in-sample, 2 years out-of-sample. The critical question for exotic signals: do they overfit more than textbook ones?

The aggregate picture (all asset classes pooled) gives one answer. But when you split by asset class, the story fractures into five very different conclusions.

Pooled slope: 0.40 (60% shrinkage). Compared to the original Atlas’s 0.93 slope, exotic signals appear to overfit dramatically. But the aggregate masks what’s really happening:

Subscribe now

Equities: 19% Shrinkage (slope 0.81)

Equity index futures are the best-behaved asset class. The IS→OOS relationship is nearly linear, with 78% sign agreement. Exotic signals on equities generalise almost as well as textbook ones — the added complexity doesn’t hurt here. Accel Fade is consistently negative in both periods, and Skew Fade 120-bar improves OOS, suggesting these effects are structural.

FX: 25% Shrinkage (slope 0.75)

G7 FX also generalises well. 61% sign agreement is lower than equities but the slope is strong. Keltner Reversion dominates both IS and OOS, which makes sense — session breaks and gaps are pronounced in FX, exactly where ATR normalization should add value. The weak spot: SkewKurt flips signs across all three horizons, suggesting higher-moment interaction signals on FX are fragile.

Bonds: 43% Shrinkage (slope 0.57)

Fixed income sits in the middle. The slope is moderate but sign agreement is the highest of any asset class at 100% — every signal that worked IS keeps the same direction OOS, just weaker. This is textbook shrinkage rather than regime change. Keltner Reversion carries the positive Sharpes; most other exotic families are negative in both periods.

Crypto: 64% Shrinkage (slope 0.36)

Crypto is the surprising case. The mean IS Sharpe is −0.35 but the mean OOS Sharpe is +0.23. Signals that looked weak or wrong in 2022–2024 started working in 2024–2026. This isn’t overfitting — it’s a regime change. Crypto’s microstructure shifted as institutional participation grew: Ret-Z Rev and Accel Fade went from negative to strongly positive. Only 44% sign agreement confirms this is a different market.

Commodities: −163% Shrinkage (slope −0.63)

Commodities are the disaster case. The regression slope is negative — a signal that worked IS actively hurts you OOS. Only 28% of signals agree on sign (worse than random). This isn’t shrinkage; this is a complete regime break. Commodity markets in 2022–2024 (post-COVID supply shocks, energy crisis) behaved fundamentally differently from 2024–2026. Exotic signals with their higher-moment dependencies amplified this instability. If you had fitted exotic strategies on commodity IS data, you would have lost money OOS.

The lesson: overfitting isn’t just a property of signal complexity — it’s an interaction between complexity and asset class stability. Exotic signals work where the underlying market dynamics are stable (equities, FX). They fail catastrophically where regimes shift (commodities). The aggregate 0.40 slope is a misleading average of a 0.81 and a −0.64.

Regime Dependence

Year-by-year Sharpes reveal whether exotic signals are regime-dependent in the same way as textbook ones. If Keltner Reversion outperforms in volatile years (when ATR is high and matters more), that’s a structural edge. If Skewness Fade only works in trending years, it’s capturing regime, not alpha.

Rolling Sharpe panels by family. Watch for whether exotic families have smoother or spikier rolling performance than textbook ones. Higher-moment signals (skewness) should be noisier; ATR-normalized signals should be smoother.

Cross-Asset Diversification

A key benefit of exotic signals: if they’re driven by different market dynamics than textbook signals, they offer additional diversification. Low cross-asset correlations combined with low textbook-vs-exotic correlations would mean a portfolio running both signal sets captures substantially more alpha than either alone.

Equity Curves by Asset Class

Cumulative risk-adjusted PnL by exotic family (6 panels) and lookback horizon (3 lines each), on 30-minute bars:

Equity Index Futures

Bond Futures

Commodity Futures

G7 FX

Crypto

Compare the slope and consistency of each exotic family’s equity curve with the corresponding textbook family from the original Atlas. Monotonically rising curves with low drawdowns indicate robust alpha. Choppy or flat curves indicate noise.

Takeaways

Exotic signals capture genuinely different effects. Skewness (third moment), acceleration (second derivative), ATR normalization (intrabar range), and return-space z-scoring each express a different view of price dynamics. They are not redundant transformations of the same signal.
ATR normalization is the lowest-hanging exotic fruit. Replacing σ with ATR in the mean reversion signal is a one-line change that incorporates OHLC information. If it improves Sharpe, adopt it unconditionally.
Higher moments need longer lookbacks. Skewness estimated from 5 bars (2.5 hours) is pure noise. The third moment needs at least 20–120 bars to stabilize. If Skew Fade 5-bar shows poor performance but 120-bar is strong, the signal is real but the estimator needs data.
Acceleration captures exhaustion patterns. The second derivative of price fires at inflection points — where a move transitions from accelerating to decelerating. This is the quantitative equivalent of ‘the trend is getting tired.’
Moment interactions unlock non-linear effects. Signed Kurtosis (skew × kurtosis) captures something neither moment captures alone: distributional extremes that are both asymmetric and fat-tailed. This is the quantitative equivalent of ‘the move looks unsustainable from every angle.’
Fractal Regime is the self-aware alpha. By measuring whether the market is geometrically trending before applying a trend signal, Fractal Trend avoids the main failure mode of momentum: whipsawing in choppy regimes. It’s regime-adaptive without a separate regime model.
The best portfolio combines textbook and exotic. If exotic signals have low correlation with textbook ones (which they should, by construction), running both families captures the full alpha landscape: textbook for robust baseline, exotic for incremental edge.

The bottom line: exotic signals are not magic. They capture genuinely different market dynamics, but whether the added complexity pays for itself depends on the specific family and asset class. The safest exotic upgrade is Keltner Reversion — same thesis as textbook MR, better normalization. Everything else is alpha research, not guaranteed improvement.

The Intraday Alpha Atlas: What Works on 30-Minute Bars — What works on 30-minute bars across asset classes
Simple Alpha Atlas: What Actually Works Across Markets (Daily) — What alpha actually works across markets on daily bars
The 7 Layers of Ensembling in Systematic Trading — Seven practical layers for combining trading strategies
5 Signal Scaling Tricks That Turn Model Predictions Into Actual Trades — From raw prediction to optimal position sizing

The Intraday Alpha Atlas: What Works on 30-Minute Bars

oracle — Wed, 15 Apr 2026 12:14:39 GMT

In the daily Alpha Atlas, commodity trend was the top alpha at Sharpe 2.20. Zoom to 30-minute bars and it flips to −2.33 — but that negative Sharpe is just as valuable. Flip the signal and you have a fade-the-breakout strategy at |Sharpe| 2.85 in bonds, 2.52 in FX, 2.33 in commodities. The alpha map doesn't just shift with frequency. It inverts — and the inversions are tradable.

This post applies the identical 12 parameter-free alphas to 30-minute bars across the same 5 asset classes. The lookback windows — 5, 20, and 120 bars — now correspond to ~2.5 hours, ~10 hours, and ~60 hours respectively. Everything else is unchanged: no optimisation, no fitting, no lookahead.

But this time, I'm adding a dimension the daily Atlas didn't have: tradability. For every signal, I compute return per trade (RPT) per unit bet — what you'd actually earn per directional trade assuming a fixed ±1 position. Intraday alpha research without a turnover lens is incomplete.

A methodological note: all signals use delay_1(close) — the previous bar's close — instead of the current close. This avoids a subtle shared-variable bias where close_t appears in both the signal and the return denominator (close_{t+1} − close_t) / close_t. Without this correction, Sharpe ratios can be inflated by 50–80%. The numbers in this post are bias-corrected.

For every strategy, I now compute: trades per year (sign-changes in the signal), average holding period in bars, and edge per trade (total PnL divided by number of directional flips). These numbers separate the alphas that are interesting on paper from those that are implementable in practice.

Also check out two similar posts about exotic alphas and also microstructure based:

The Exotic Alpha Atlas: Beyond Textbook Signals on 30-Minute Bars

Crypto High-Frequency Simple Alpha Atlas

Signal Definitions

12 alphas from 4 families, each computed at three lookback horizons N ∈ {5, 20, 120} bars:

Each family is oriented so that positive Sharpe = winning direction on intraday bars. Families where the original trend/momentum direction loses intraday are flipped to their profitable mean-reversion counterpart:

Fade Breakout. Go short when P_t > max(P_{t-N:t-1}), long when P_t < min(P_{t-N:t-1}). Fades Donchian channel breakouts — the opposite of trend-following. At N=5 this fades breakouts of the last 2.5 hours.
Fade MA. Go short when price is above the N-bar SMA, long when below. Fades MA crossovers. At N=120, this fades a 2.5-day moving average.
CS Reversion. Buy relative losers, sell relative winners — the negative of cross-sectional momentum. Rank assets by P_t / SMA_N, go short the strongest and long the weakest.
Mean Reversion (MR). s_t = −z_t where z_t = (P_t − μ_N) / σ_N. Buy oversold, sell overbought. The classic time-series mean-reversion alpha.

Applied identically to 29 instruments across 5 asset classes:

Equity index futures: ES, NQ, YM, NKD, FTSE, DAX
Bond futures: TN, UB, ZN, ZF, ZT
Commodity futures: GC, SI, CL, HG, ZW, ZC
G7 FX: EUR, GBP, JPY, AUD, CAD, CHF vs USD
Crypto: BTC, ETH, SOL, XRP, DOGE, ADA

Period: March 2022 – February 2026 (~4 years). Sharpe ratios annualised from monthly returns. Each instrument contributes ~50,000 bars — roughly 40× the sample size of a daily backtest over the same period.

The Alpha × Asset Class Heatmap

Each cell shows the annualised Sharpe (left) and return per trade per unit bet (right) for one alpha–asset class pair:

Left: annualised Sharpe. Right: RPT — return per trade assuming a ±1 unit position. All signals use delay_1(close). No costs, no fitting, no lookahead.

Four patterns jump out:

Fade-the-breakout dominates. Fade 5-bar breakout is the top alpha in bonds (2.85), FX (2.51), commodities (2.33), and crypto (1.79). This is the single strongest finding: intraday breakouts don't follow through — they revert.
All four families are now mean-reverting. Fade breakout, Fade MA, CS reversion, and time-series MR all express the same underlying effect: intraday price overshoots revert. Bonds and FX are 12/12 positive. Commodities 10/12. The heatmap is predominantly green.
Fade breakout beats continuous MR. Fade 5-bar breakout at Sharpe 2.3–2.9 roughly doubles MR 5-bar zscore at 1.0. The ternary signal (fire only at range extremes) is structurally cleaner than the continuous z-score approach.
Equities remain a coin flip. Best equity alpha: CS-Rev 5-bar at 0.47. Only 3/12 strategies are positive. Equity index futures are too efficiently arbitraged for simple signals to gain traction at any frequency.

Best Strategy per Asset Class

When we treat negative Sharpes as equally tradable (by flipping the signal), the picture is clear: flipped 5-bar breakout is the strongest signal in 4 of 5 classes at |Sharpe| 1.8–2.9. MR 5-bar zscore is the strongest unflipped signal in bonds (1.04), FX (1.03), and commodities (0.96).

Both are expressing the same insight: at intraday frequencies, the dominant regularity across liquid markets is transient mean reversion. The flipped breakout implements this by fading range extremes; the MR z-score implements it continuously. The breakout version is stronger because it fires selectively at extreme points rather than continuously.

The Overfitting Tax

I split the data at March 2024 — 2 years in-sample, 2 years out-of-sample — and regressed OOS Sharpe on IS Sharpe for every alpha–asset combination. All signals are already oriented in the profitable direction.

The core advantage of intraday data: ~40× more observations per unit time means your Sharpe estimates are structurally more precise than daily backtests. What you see in-sample is closer to what you get out-of-sample.

The Tradability Lens: Turnover, Holding Period, and Edge per Trade

Gross Sharpe is necessary but not sufficient. Every strategy here was evaluated for three additional metrics computed directly from the signals:

Trades per year — the number of directional sign-changes in the signal (a position flip from long to short, or vice versa).
Average holding period — mean number of 30-minute bars between sign-changes. A holding period of 1.3 bars means the signal flips direction roughly every 40 minutes.
Edge per trade — total PnL divided by the number of sign-changes. This is what survives (or doesn't) after you pay the spread.

Here are the top 10 positive-Sharpe strategies with their tradability profiles:

All signals oriented so positive Sharpe = winning direction. Fade = short on breakout/MA crossover (mean reversion at extremes). CS-Rev = buy cross-sectional losers, sell winners. RPT = return per trade per unit bet. Mar 2022 – Feb 2026.

A signal with Sharpe −2.85 is exactly as tradable as one with +2.85 — you just take the opposite position. The table above ranks by |Sharpe| and marks flipped signals with *. Key patterns:

Fade-the-breakout dominates. Flipping the 5-bar breakout signal (go short when price breaks above the range, long when it breaks below) produces |Sharpe| 2.3–2.9 in bonds, FX, and commodities. This is the strongest intraday signal in the Atlas — a mean-reversion strategy implemented at range extremes.
All top signals are mean-reverting. Whether implemented as flipped breakouts (ternary, firing at extremes) or MR z-scores (continuous), the winning edge is the same: intraday price overshoots revert. The breakout version is stronger because it only fires at extreme points and has cleaner signal construction.
RPT ranges from 0.29 to 6.64. Higher RPT means more room to absorb execution costs. Crypto trend 120-bar has the best RPT (6.64) thanks to higher volatility, but wider spreads and slippage partially offset this.
5-bar signals trade 9,000–16,000 times/year with 1-bar holding periods. The 120-bar signals trade 2,000–4,000 times/year with 4–5 bar holds — better tradability profiles for markets with wider spreads.

Regime Dependence

Year-by-year Sharpes for the best strategy per asset class show regime dependence similar to daily. Note that intraday alphas can pick up regime transitions faster — a macro shock that takes days to propagate through close-to-close returns appears within hours on 30-minute bars.

Rolling Sharpe panels are notably smoother than their daily counterparts — an artefact of ~40× more observations per window. This makes conditional performance evaluation more reliable: you can distinguish a genuine regime shift from sampling noise with much less lag.

Diversification at Higher Frequency

Intraday cross-asset correlations are structurally lower than daily. At the 30-minute scale, asset-specific microstructure effects — session times, venue fragmentation, market-maker inventory cycles — dominate over shared macro factors. The implication: multi-asset deployment offers even greater diversification benefit intraday than daily.

Equity Curves by Asset Class

Cumulative risk-adjusted PnL by alpha family (4 panels) and lookback horizon (3 lines each), on 30-minute bars:

Equity Index Futures

Bond Futures

Commodity Futures

G7 FX

Crypto

Note the MR-5 curves in bonds, FX, and commodities: steep, consistent, and almost monotonically increasing. Contrast with the CS panels: uniformly negative across all asset classes.

Takeaways

The alpha map inverts — and the inversions are tradable. Daily commodity trend (+2.20) becomes −2.33 on 30M. Flip that signal and you have a fade-the-breakout strategy at |Sharpe| 2.33. Negative Sharpes are just as valuable as positive ones.
Fade-the-breakout is the strongest intraday alpha. Flipped 5-bar breakout achieves |Sharpe| 2.3–2.9 in bonds, FX, and commodities. It's a mean-reversion strategy that fires selectively at range extremes — stronger and cleaner than continuous z-score approaches.
RPT separates tradable from theoretical. Flipped breakout in bonds has RPT 0.85 across ~9,000 trades/year. Crypto trend 120-bar has RPT 6.64 across ~3,800 trades. Higher RPT means more room for execution costs.
Everything is mean reversion at 30 minutes. Whether via flipped breakouts, z-score fading, or flipped MA crossovers, every winning strategy exploits the same microstructure effect: intraday overshoots revert within 1–2 bars.
The strongest portfolio runs both timescales. Daily trend + intraday fade-the-breakout: structurally uncorrelated, mechanistically distinct, and mutually beneficial for execution.

The bottom line: zoom from daily to 30-minute and the alpha map inverts. The strongest intraday signal is fade-the-breakout — flipping the very trend signals that work on daily bars. The best systematic portfolio runs both timescales, capturing daily trends and intraday reversions as complementary, uncorrelated legs.

The Exotic Alpha Atlas: Beyond Textbook Signals on 30-Minute Bars — Beyond-textbook alpha signals on 30-minute bars
Simple Alpha Atlas: What Actually Works Across Markets (Daily) — What alpha actually works across markets on daily bars
5 Signal Scaling Tricks That Turn Model Predictions Into Actual Trades — From raw prediction to optimal position sizing
The 7 Layers of Ensembling in Systematic Trading — Seven practical layers for combining trading strategies

Monthly Alpha Hunt: March 2026

oracle — Tue, 14 Apr 2026 16:05:46 GMT

The idea is simple: every month I generate a huge number of random mathematical expressions, evaluate each one as a trading signal on historical crypto data, and see what sticks. No optimization, no gradient descent — pure random search. This month's standout is a formula on 5M bars that delivered a 2.88 Sharpe over two years. The full report covers Daily, 30M, and 5M results with equity curves.

All genes are evaluated on a 2-year rolling window (Mar 2024 – Mar 2026) on a crypto assets basket.

Methodology: Each gene provides the entry signal. Positions are sized by prediction magnitude and held on a rolling 5-bar window — at any point, the portfolio position reflects the sum of the last 5 predictions. Returns are realized continuously across this window.

Executive Summary

Daily Timeframe

Risk-Adjusted Equity Curves

Performance comparison with volatility normalized to 15% annualized:

30M Timeframe

Risk-Adjusted Equity Curves

Performance comparison with volatility normalized to 15% annualized:

5M Timeframe

Risk-Adjusted Equity Curves

Performance comparison with volatility normalized to 15% annualized:

Recommendation

The strongest gene was found on the 5M timeframe with an annualized Sharpe of 2.878:

tsrank(17, plus(rsi(12)))

How to Read the Genes

Each gene is a formula assembled from simple building blocks. Here is what each function does:

rsi(N) — Relative Strength Index with N-bar lookback. Classic momentum oscillator (0–100).

lag(N, x) — The value of x from N bars ago.

sign(x) — Returns +1 if positive, −1 if negative.

sum(N, x) / mean(N, x) / median(N, x) — Rolling sum, average, or median over the last N bars.

cor(N, x, y) — Rolling Pearson correlation between x and y over N bars.

kentau(N, x, y) — Kendall tau rank correlation over N bars. Measures monotonic trend agreement.

tsrank(N, x) — Time-series percentile rank: where the current value sits within its own N-bar history.

crossrank(N, x) — Cross-sectional rank: rank of each asset relative to the others at each timestamp.

tsmin(N) / tsmax(N) — Rolling min / max over N bars.

plus, minus, mult, div — Standard arithmetic.

neg(x) / abs(x) — Negate or take absolute value.

This Month's Winners — In Plain English

Daily #1 (Sharpe 1.56): log(sum(2, plus(rsi(30))))

Accumulates the sign or magnitude of RSI over a rolling window — trades on sustained momentum rather than single-bar readings.

Daily #2 (Sharpe 1.56): log(plus(rsi(14)))

Uses RSI momentum as the primary signal — goes long when momentum is elevated, short when it is depressed.

30M #1 (Sharpe 1.52): tsmax(12, cor(27, tsmax(9, rsi(39)),minus(rsi(16))))

Subtracts RSI extremes from smoothed price levels, creating a mean-reversion signal that fires when momentum overshoots.

30M #2 (Sharpe 1.46): kentau(21, lag(3, sum(19, rsi(29))),mult(tsmin(18),sign(rsi(24))))

Measures the rank correlation between RSI momentum and price over a rolling window. When momentum and price trend together, the signal strengthens; divergences flip it.

5M #1 (Sharpe 2.88): tsrank(17, plus(rsi(12)))

Uses RSI momentum as the primary signal — goes long when momentum is elevated, short when it is depressed.

5M #2 (Sharpe 2.81): mean(12, cor(31))

Smooths the rolling auto-correlation of price — a persistence detector. High values mean price is trending; low values mean it is choppy or reversing.

Disclaimer: This report is for informational and educational purposes only. It does not constitute financial advice, investment recommendations, or a solicitation to buy or sell any securities or crypto assets. All results are based on historical backtests with simulated execution — past performance is not indicative of future results. Trading involves substantial risk of loss. Always do your own research and consult a qualified financial advisor before making any investment decisions.

Monthly Alpha Hunt: Millions of Random Guesses, One Good Trade — Millions of random expressions, one good trade
Monthly Crypto Alpha Hunt: Millions of Random Guesses, One Good Trade (Feb 2024-Jan 2026) — The first monthly alpha hunt across timeframes
The Exotic Alpha Atlas: Beyond Textbook Signals on 30-Minute Bars — Beyond-textbook alpha signals on 30-minute bars
5 Signal Scaling Tricks That Turn Model Predictions Into Actual Trades — From raw prediction to optimal position sizing

Crypto Strategy League Tables: March 2026

oracle — Sat, 04 Apr 2026 11:51:05 GMT

Every month, I test four classic trading strategies with identical parameters: trend breakouts, moving average crossovers, cross-sectional momentum, and mean reversion. Same rules, same coins, same timeframes. No optimization, no cherry-picking.

March delivered the most dramatic regime shift in the two-year sample. On daily bars, trend-following — the dominant strategy family for two years — collapsed. Mean reversion posted its best single month ever. On 5-minute data, the momentum machine barely flinched. And on 30-minute bars, mean reversion quietly consolidated its new position at the top of the leaderboard.

March’s Winners

The same month, the same coins — but three different winners depending on your timeframe. This is the recurring theme of this report, and March makes the case more forcefully than ever.

The 4 Strategy Types

TR-BO (Breakout): Long above N-bar high, short below N-bar low
TR-MA (Moving Average): Long above the N-bar moving average, short below
CS (Cross-Sectional Momentum): Rank coins by relative strength; buy leaders, sell laggards
MR (Mean Reversion): Z-score based; buy oversold, short overbought

All tested with lookback windows of 5, 20, and 120 bars across 19 major cryptocurrencies. All strategies are long/short.

Daily Strategies

March was a bloodbath for daily trend-following. Every short-lookback trend strategy — TS 5-day MA, TS 20-day MA, TS 5-day breakout — posted its worst risk-adjusted month of the two-year sample.

The sole trend survivor? TS 120-day MA, the only trend strategy to finish March positive — confirming that longer horizons provide crucial insulation against short-term reversals.

But the real story is mean reversion. MR 5-day and MR 20-day z-score both posted their strongest risk-adjusted months of the entire two-year dataset. After February’s MR surge on 30-minute bars, March brought the same regime to daily data.

Despite the March fireworks, mean reversion remains uniformly negative on daily data over the full two-year period, with Sharpes from -0.17 to -0.50. One month — even a spectacular one — doesn’t erase 23 months of losses.

Takeaway: the trend regime is cracking on daily data. Two years of dominance masked an important truth — when trend reverses, it reverses violently. Position sizing and lookback diversification matter more than ever.

30-Minute Strategies

The 30-minute timeframe continues its quiet revolution. MR 120-bar z-score leads the leaderboard at 0.32 Sharpe, up from 0.30 last month. CS Z-Score 120-bar follows at 0.24, and CS Z-Score 20-bar at 0.21.

March was modest but positive for the leaders — no fireworks, just steady risk-adjusted accumulation. What started as a two-month anomaly in January-February is now a three-month trend. At what point does “anomaly” become “regime”?

The paradox deepens: mean reversion now leads at 30-minute frequency and had its best month ever on daily data. But the mechanisms are different. Daily MR captures multi-day overshoots driven by macro shocks. 30-minute MR captures intraday mean reversion within established ranges. Same label, different dynamics.

5-Minute Strategies

1.21 Sharpe — CS Z-Score 120-bar on 5-minute data — down from 1.25 last month.

For the first time in six months, the 5-minute momentum machine decelerated. CS Z-Score 120-bar dipped from 1.25 to 1.21 Sharpe, with a flat March. CS Z-Score 20-bar fared better, maintaining its 0.89 Sharpe.

Still institutional-grade. Still the best risk-adjusted strategy in the entire study. But the March regime shift — which devastated daily trend and energized daily mean reversion — was invisible at 5-minute frequency. The intraday momentum signal operates on a different clock.

Mean reversion remains consistently negative on 5-minute data, with Sharpes from -0.20 to -0.72. The high-frequency regime is unambiguously momentum-driven.

The Timeframe Paradox

March makes the case more starkly than any previous month:

MR 5-day z-score posted its best risk-adjusted month ever on daily data. On 5-minute data, the same strategy has a cumulative Sharpe of -0.20.
CS Z-Score 120-bar compounded at 1.21 Sharpe on 5-minute data. On daily data, its Sharpe is 0.32.
Trend-following collapsed on daily bars in March but was flat-to-positive on 5-minute bars.

Different frequencies capture different market dynamics. Daily bars reflect macro regime shifts — the March selloff, rotation, and short squeeze. Five-minute bars capture intraday flow rotation across coins. Thirty-minute bars sit at the boundary, catching intraday reversions. Neither timeframe is “right.” All are real. The edge is in knowing when each applies — or better, running all three.

Performance by Asset

Daily

30-Minute

5-Minute

The 2-Year View: Who Survives?

Daily — Trend Still Rules, but Cracks Appeared

TS 120-day breakout leads at 0.61 Sharpe, down from 0.72 last month — the sharpest single-month decline in the series. March’s trend reversal pulled every trend strategy lower. The TS 120-day MA sits at 0.44, the TS 5-day breakout at 0.35. Over two years on daily bars, fighting the trend is still a losing proposition — but the margin of safety is narrowing.

5-Minute — Momentum’s Kingdom

1.21 Sharpe — CS Z-Score 120-bar on 5-minute data — the best risk-adjusted strategy in the entire study.

Down from 1.25 but still dominant. CS Z-Score 20-bar follows at 0.89. TS 120-bar MA holds at 0.71. The 5-minute timeframe continues to be the most rewarding frequency for systematic strategies. Mean reversion remains a consistent money-loser at this frequency.

30-Minute — Mean Reversion Consolidates

MR 120-bar now leads at 0.32 Sharpe, extending its lead over CS Z-Score 120-bar (0.24). Three consecutive months of sustained MR outperformance have reshaped the full-period statistics. What looked like a one-off in January is now the longest MR winning streak in the sample.

6-Month Rolling History

What to Watch in April

Was March’s daily MR explosion a one-off or a regime shift? MR 5-day z-score posted its best risk-adjusted month ever. If April confirms with another positive MR month on daily data, we may need to reassess the “trend always wins on daily” assumption that has held for two years.
The 30-minute MR streak is now three months. January, February, March — MR 120-bar has led the 30-minute leaderboard every month. Is this structural or cyclical? The answer determines whether multi-timeframe ensembles should permanently allocate to 30M MR.
5-minute momentum decelerated for the first time. CS Z-Score 120-bar went from 1.13 → 1.25 → 1.21 Sharpe. One down-tick is noise. Two would be a signal. Watch whether March’s daily regime turbulence begins bleeding into intraday momentum.

Data period: February 2024 – March 2026. Universe: BTC, ETH, SOL, ADA, DOGE, XRP, LTC, DOT, LINK, AVAX, TRX, and others. All strategies are long/short.

Disclaimer: Educational and research purposes only. Does not constitute investment advice. Past performance is not indicative of future results. Strategies do not include transaction costs, slippage, or market impact.

Crypto Price Action Alpha Report: February 2026 — Crypto strategy performance, February 2026
Crypto - Price Action Alpha Report - Jan 2026 — Crypto strategy performance, January 2026
The Replicator: Catching Crypto Trends — Replicating a Donchian trend strategy for crypto
Simple Alpha Atlas: What Actually Works Across Markets (Daily) — What alpha actually works across markets on daily bars

Simple Alpha Atlas: What Actually Works Across Markets (Daily)

oracle — Fri, 20 Mar 2026 14:44:09 GMT

Trend-following works in commodities. Mean reversion works in equities. Everyone knows this.

But does everyone test it? I did. 12 alphas, 5 asset classes, 5 years, zero optimization, and no transaction costs. Here’s what the data actually says — and what you can do with it.

The Setup

12 alphas across 4 signal families, each with 3 lookback horizons (5, 20, 120 days):

Trend Breakout (TR-BO): Long above N-day high, short below N-day low.
Trend Moving Average (TR-MA): Long above N-day SMA, short below.
Cross-Sectional Momentum (CS): Rank by relative strength, bet the spread.
Mean Reversion (MR): Fade the Z-score. Buy oversold, sell overbought.

5 asset classes, 29 instruments:

Equity Index Futures: ES, NQ, YM, NKD, FTSE, DAX
Bond Futures: TN, UB, US, ZN, ZF
Commodity Futures: Gold, Silver, Crude, Copper, Wheat, Corn
G7 FX: EUR, GBP, JPY, AUD, CAD, CHF vs USD
Crypto: BTC, ETH, SOL, XRP, DOGE, ADA

Period: March 2021 to February 2026. Same parameters everywhere. No fitting. 60 alpha–asset combinations.

The Heatmap: Where Alpha Lives

Read this chart carefully — it’s the most information-dense image in this post. Each cell is an annualized Sharpe ratio. Green = money. Red = pain. But red is also money: a Sharpe of −1.44 means the reverse trade has a Sharpe of +1.44. Every deep-red cell is a strong alpha hiding in plain sight — you just need to flip the sign.

FX mean reversion dominates. The 5-day Z-score hits Sharpe 1.29. But look at the 5-day MA: −1.44. That’s the same edge seen from the other side — trend-following fails in FX because FX mean-reverts. Flip the trend signal and you get a second alpha with an even higher Sharpe.
The heatmap is symmetric in disguise. Cross-sectional momentum in FX: −1.40. Flip it (buy relative losers, sell winners) and you have a Sharpe 1.40 contrarian alpha. The “worst” cells are the best trades — you’re just looking at them backwards.
Crypto rewards patience. 120-day breakout at 0.88. Short-term crypto signals are noise. The money is in catching regime shifts and holding.
Equities revert. 120-day MR at 0.63. Trend-following equity indices is a losing game over this period — but that means fading the trend is a winning one.
Commodity trend is real but thin. Best Sharpe 0.53. It works — just not as dramatically as the CTA marketing suggests.
Bonds: almost nothing works. Sharpes cluster near zero in both directions. There’s no strong alpha to flip either — bonds are just hard.

Action point: Don’t skip the red cells. A Sharpe of −1.44 on FX trend-following is a Sharpe of +1.44 on FX mean-reversion. The heatmap gives you 60 alphas, not 60 — because every negative cell is a positive alpha with the sign flipped. Scan for the deepest reds. Those are your strongest signals.

Why These Alphas Work (and Don’t)

Grinold’s Fundamental Law: IR ≈ IC × √Breadth. IC is how right your signal is per trade. Breadth is how many trades you make per year. You need both.

FX MR (Sharpe 1.29): IC ≈ 0.05 — the highest in the study. FX markets are liquid, mean-reverting at short horizons, and the IC barely fluctuates year-to-year. This is a structurally stable edge.
Crypto Breakout (0.88): IC ≈ 0.03. Moderate signal, but crypto’s tendency to trend for months at a time makes even a weak IC compound into a decent Sharpe.
Equity MR (0.63): IC ≈ 0.025. Institutional rebalancing creates predictable countertrend flows in index futures. Slow but reliable.
Commodity Trend (0.53): IC ≈ 0.02. Weak per-bet. The edge is real — physical supply/demand imbalances do create trends — but you need a broad commodity universe to make the Fundamental Law work in your favor.
Bond Trend (0.51): IC ≈ 0.02, and unstable. Rate cycles create episodic trends that look systematic in backtests but don’t persist.

Action point: If your commodity trend alpha disappoints, the fix isn’t a better signal — it’s more instruments. The IC is weak but positive. Adding breadth (more commodity contracts, more commodity-like markets) is the highest-leverage improvement you can make.

The Cheat Sheet: Best Alpha per Market

Cut to the chase:

FX: 5-day mean reversion (Sharpe 1.29)
Crypto: 120-day breakout (Sharpe 0.88)
Equities: 120-day mean reversion (Sharpe 0.63)
Commodities: 120-day breakout (Sharpe 0.53)
Bonds: 20-day MA (Sharpe 0.51)

Notice the pattern: liquid, information-rich markets (FX, equities) revert. Physical and momentum-driven markets (commodities, crypto) trend. This isn’t a coincidence — it’s microstructure. And remember: if a market strongly trends, the reverse-trend signal (mean reversion) will show a large negative Sharpe — which is just a large positive Sharpe for the opposite trade. The best alphas often hide in the worst-looking cells.

Action point: Pick the alpha family that matches your market’s microstructure, not the one you wish worked. If you trade FX and you’re running trend-following, you’re fighting the market’s natural behavior.

The Regime Problem

Crypto 2022–2023: Sharpes of 1.71 and 1.09. The crash-then-recovery was a textbook trending environment.
Equities: The most consistent — positive in 4 of 6 years, peaking at 1.89 in early 2026.
FX: Swings from −1.42 (2025) to strongly positive in other years. High alpha, high variance.
Commodities: Negative in several years. Not the all-weather edge it’s marketed as.

The year-to-year variance within each asset class dwarfs the differences between them. The best market in one year can be the worst the next.

Action point: Never allocate 100% to one market based on a backtest. The Sharpe you see is an average over regimes. In any given year, that average is meaningless. Spread your capital across asset classes and alpha families — the diversification is the alpha.

Rolling Sharpe: How Stable Is the Edge?

These panels answer the question every trader asks: is the alpha dying? FX mean reversion (red) stays consistently above zero. Crypto (purple) spikes and fades. No single market or alpha family delivers a steady paycheck — which is why you need all of them running simultaneously.

Cross-Asset Correlations

Low to moderate correlations across the board. Equity and bond returns move together (macro sensitivity). Crypto is its own animal. This confirms that multi-asset deployment isn’t just a good idea — it’s free risk-adjusted return.

Per-Class Equity Curves

Risk-adjusted equity curves by alpha family (5d, 20d, 120d lookbacks):

Equity Index Futures

Bond Futures

Commodity Futures

FX (G7)

Crypto

The Overfitting Tax: How Much of This Is Real?

Everything above uses the full sample. But the only question that matters: how much survives out of sample?

I split at March 2024 — 3 years in-sample, 2 years out-of-sample — and compared every alpha’s Sharpe across the two periods.

The dashed line is “what you backtest = what you get.” The red line is reality.

Slope: 0.24. For every 1.0 Sharpe in your backtest, expect to keep 0.24. 76% of your backtest Sharpe is noise.

FX (red): Most coherent IS→OOS. Mean-reversion alphas persist. This is the most “real” edge in the study.
Crypto (purple): Mixed. A few breakout signals survive; many flip sign.
Equities (blue): Honest but uninspiring. IS near zero predicts OOS near zero.
Commodities (green): Concerning — several positive IS Sharpes go negative OOS.
Bonds (orange): Worst offenders. The rate-hiking cycle looked like a trend. It wasn’t.

Action point: Haircut every backtest Sharpe by 75% before you size a position. If it’s not positive after the haircut, don’t trade it. If a signal requires optimization to show an edge, the edge is probably in the optimization, not the market.

What to Do With This

This study gives you a map. Here’s how to use it:

Start with the highest-IC market for your signal type. If you’re building mean-reversion alphas, start with FX. If you’re building trend alphas, start with crypto or commodities. Don’t try to force a signal type into a market that doesn’t support it.
Use the Fundamental Law to diagnose underperformance. If your Sharpe is lower than expected, decompose it. Is the IC too low (bad signal)? Or is the breadth too low (too few instruments)? The fix is completely different in each case.
Build multi-asset, multi-alpha portfolios. The correlations are low. Running all 12 alphas across all 5 asset classes gives you ~60 bets with moderate IC and massive breadth. The portfolio Sharpe will be substantially higher than any individual cell in the heatmap.
Haircut everything by 75%. Your backtest Sharpe is not your live Sharpe. Size positions to the post-haircut number. If a strategy needs leverage to work after the haircut, it doesn’t work.
Diversify across time horizons. The 5-day, 20-day, and 120-day alphas have different regime sensitivities. Running all three is cheap insurance against any one horizon going cold.
Rerun this on your own data. The methodology is deliberately simple. You can replicate this with any set of instruments, any lookback periods, any time range. The value isn’t in my specific numbers — it’s in the framework. Build your own heatmap. Find your own alpha atlas.

Key Takeaways

There is no universal alpha. Match the signal family to the market’s microstructure.
FX mean reversion is the best simple alpha in this study. Sharpe 1.29 from a 5-day Z-score. Highest IC, most stable OOS.
Commodity trend is overrated as a standalone. Sharpe 0.53 with weak IC. The edge is real but needs breadth to compound.
76% of your backtest is noise. Haircut aggressively. Size conservatively.
The portfolio is the product. No single alpha or market is the answer. Diversification across both is where the Sharpe lives.

The alpha atlas doesn’t show you one treasure. It shows you that the treasure is the map itself.

Disclaimer: This post is for educational and informational purposes only. Nothing here constitutes financial advice, investment advice, or a recommendation to buy, sell, or hold any security, derivative, or cryptocurrency. The analysis uses historical data with no transaction costs — actual trading results will differ materially due to slippage, commissions, market impact, and execution risk. Past performance does not predict future results. All Sharpe ratios, returns, and statistics are hypothetical and based on backtested data that was not traded live. Do your own research, understand your risk tolerance, and consult a qualified financial advisor before making any investment decisions. The author may hold positions in the instruments discussed.

The Intraday Alpha Atlas: What Works on 30-Minute Bars — What works on 30-minute bars across asset classes
The Exotic Alpha Atlas: Beyond Textbook Signals on 30-Minute Bars — Beyond-textbook alpha signals on 30-minute bars
The 7 Layers of Ensembling in Systematic Trading — Seven practical layers for combining trading strategies
5 Signal Scaling Tricks That Turn Model Predictions Into Actual Trades — From raw prediction to optimal position sizing

Monthly Alpha Hunt: Millions of Random Guesses, One Good Trade

oracle — Wed, 04 Mar 2026 09:09:52 GMT

The idea is simple: every month I generate a huge number of random mathematical expressions, evaluate each one as a trading signal on historical crypto data, and see what sticks. No optimization, no gradient descent — pure random search. This month's standout is a formula on 5M bars that delivered a 2.81 Sharpe over two years. The full report covers Daily, 30M, and 5M results with equity curves.

All genes are evaluated on a 2-year rolling window (Feb 2024 – Feb 2026).

Subscribe now

Executive Summary

Daily Timeframe

Risk-Adjusted Equity Curves

Performance comparison with volatility normalized to 15% annualized:

30M Timeframe

Risk-Adjusted Equity Curves

Performance comparison with volatility normalized to 15% annualized:

5M Timeframe

Risk-Adjusted Equity Curves

Performance comparison with volatility normalized to 15% annualized:

Recommendation

The strongest gene was found on the 5M timeframe with an annualized Sharpe of 2.808:

mean(12, cor(31))

How to Read the Genes

Each gene is a formula assembled from simple building blocks. Here is what each function does:

rsi(N) — Relative Strength Index with N-bar lookback. Classic momentum oscillator (0–100).

lag(N, x) — The value of x from N bars ago.

sign(x) — Returns +1 if positive, −1 if negative.

sum(N, x) / mean(N, x) / median(N, x) — Rolling sum, average, or median over the last N bars.

cor(N, x, y) — Rolling Pearson correlation between x and y over N bars.

kentau(N, x, y) — Kendall tau rank correlation over N bars. Measures monotonic trend agreement.

tsrank(N, x) — Time-series percentile rank: where the current value sits within its own N-bar history.

crossrank(N, x) — Cross-sectional rank: rank of each asset relative to the others at each timestamp.

tsmin(N) / tsmax(N) — Rolling min / max over N bars.

plus, minus, mult, div — Standard arithmetic.

neg(x) / abs(x) — Negate or take absolute value.

This Month's Winners — In Plain English

Daily #1 (Sharpe 1.42): plus(lag(3, sign(rsi(37))),sum(3, lag(9)))

Takes the direction (+1 or −1) of RSI momentum from several bars ago and combines it with lagged prices — a delayed momentum-direction signal.

Daily #2 (Sharpe 1.40): plus(rsi(23))

Uses RSI momentum as the primary signal — goes long when momentum is elevated, short when it is depressed.

30M #1 (Sharpe 1.12): kentau(12, rsi(13))

Measures the rank correlation between RSI momentum and price over a rolling window. When momentum and price trend together, the signal strengthens; divergences flip it.

30M #2 (Sharpe 1.11): mean(19, rsi(29))

Accumulates the sign or magnitude of RSI over a rolling window — trades on sustained momentum rather than single-bar readings.

5M #1 (Sharpe 2.81): mean(12, cor(31))

Smooths the rolling auto-correlation of price — a persistence detector. High values mean price is trending; low values mean it is choppy or reversing.

5M #2 (Sharpe 2.73): neg(plus(cov(33,rsi(30)),neg(rsi(25))))

Uses RSI momentum as the primary signal — goes long when momentum is elevated, short when it is depressed.

Crypto - Price Action Alpha Report - Jan 2026 — Data-driven analysis of cryptocurrency price patterns
5 Signal Scaling Tricks That Turn Model Predictions Into Actual Trades — From raw prediction to optimal position sizing
The 7 Layers of Ensembling in Systematic Trading — Seven practical techniques for combining trading strategies

Crypto Price Action Alpha Report: February 2026

oracle — Tue, 03 Mar 2026 18:55:59 GMT

Every month, we put four classic trading strategies through the same test: trend breakouts, moving average crossovers, cross-sectional momentum, and mean reversion. Same rules, same coins, same timeframes. No optimization, no cherry-picking.

The goal is simple: find what actually works in crypto — and what doesn't. February's selloff reshuffled the deck: trend-following held firm on daily bars, mean reversion posted its best month ever on 30-minute bars, and CS momentum kept compounding on 5-minute data. Three timeframes, three different winners — again.

February's Winners

Cut through the noise. Here's what worked best at each timeframe:

The February crypto selloff produced the largest single-month MR gains in the entire two-year sample. But the winners vary by timeframe — trend-following held firm on daily bars while mean reversion swept the 30-minute podium.

How to Read This: February was defined by a broad selloff. Daily rewarded trend-followers who held through the chaos (long-horizon breakouts thrived). 30-minute rewarded contrarians (fade intraday moves — MR posted its best month ever). 5-minute rewarded speed (short-horizon trend and mixed signals). The lesson: the same month, the same coins — but three different winners depending on your timeframe.

The 4 Strategy Types

TR-BO (Breakout) — Long above N-bar high, short below N-bar low. The classic channel breakout.
TR-MA (Moving Average) — Long above the N-bar moving average, short below. Simple, old-school, effective.
CS (Cross-Sectional Momentum) — Rank all coins by relative strength. Buy the leaders, sell the laggards.
MR (Mean Reversion) — Z-Score based. Buy the oversold, short the overbought. The contrarian's bet.

Lookback windows: 5, 20, and 120 bars. Universe: 19 major cryptocurrencies. All strategies are long/short.

February 2026: The Scoreboard

Daily Strategies

0.72 Sharpe — TS 120-day breakout leads the daily leaderboard

Longer-horizon breakout strategies thrived during February's turbulence. The 120-day breakout rose to the top of the daily table (0.72 Sharpe), displacing the 5-day breakout (now 0.57). February's monthly numbers: TS 120-day breakout posted +7,048, while MR 5-day surged +13,094 — but the other MR strategies lost heavily (MR 20-day at -11,598, MR 120-day at -8,986). An unusual split.

Mean reversion remains uniformly negative on daily data over the full two-year period, with Sharpes from -0.31 to -0.55. When crypto trends, the contrarians still pay dearly at this timeframe.

The takeaway: Daily trend-following still works, but the edge has shifted toward longer horizons. In a more volatile environment, 120-day lookback provides better signal quality than 5-day.

30-Minute Strategies

This is where the story has changed dramatically.

Mean reversion now sits at the top of the 30-minute leaderboard. MR 120-bar zscore leads at 0.30 Sharpe, followed by CS Z-Score 120-bar at 0.22 and MR 20-bar at 0.20. February's monthly numbers were staggering: MR 20-bar posted +20,092, MR 5-bar at +17,324, MR 120-bar at +13,750. The single best month for mean reversion in the entire two-year sample.

The paradox: What looked like a one-off in January is now a two-month trend. The 30-minute timeframe is no longer "hostile to everything except CS momentum" — it's now hostile to trend-following specifically, while MR and CS both produce positive Sharpes. This is a meaningful update to the prior narrative.

5-Minute Strategies

The 5-minute timeframe remains the most strategy-friendly environment.

1.25 Sharpe — CS Z-Score 120-bar on 5-minute data — up from 1.13 last month

CS Z-Score 120-bar holds its crown and actually improved from last month's 1.13. Trend strategies contribute positively across most horizons, with the 120-bar MA joining the 0.85 Sharpe club. Mean reversion is the only family that consistently fails here, with all three MR strategies posting negative Sharpes. February was mixed-to-positive: short-horizon trend strategies fared best, while the 5-minute timeframe continues to be the least regime-dependent of the three.

The Timeframe Paradox

Here's the part that should make you rethink everything you know about "what works in crypto":

MR 20-bar posted +20,092 on 30-minute data in February. On daily data in the same month, MR 20-day posted -11,598. The same strategy family, the same assets, the same calendar month — opposite results.

This isn't noise. It reflects fundamentally different dynamics at different frequencies. Daily bars capture macro regime shifts — rate decisions, regulatory news, ETF flows. 30-minute bars capture intraday mean reversion — the market's tendency to overshoot and snap back within the same session.

Neither timeframe is "right." Both are real. The edge is in knowing when each applies.

Performance by Asset

Which coins rewarded which strategies? The asset-level tables show the best-performing horizon for each strategy type per coin.

Daily

30-Minute

5-Minute

The 2-Year View: Who Survives?

Monthly returns are exciting. But the real question is: which strategies hold up over two years? These are risk-adjusted returns (Sharpe ratios) since February 2024.

Daily — Trend Rules the Long Game

TR-BO 120-day leads the daily leaderboard at 0.72 Sharpe, up from 0.63 last month. All top-ranked strategies are either trend or momentum. Over two years, on daily bars, fighting the trend is still a losing proposition.

5-Minute — Momentum's Kingdom

1.25 Sharpe — CS Z-Score 120-bar on 5-minute data — the best risk-adjusted strategy in the entire study

Institutional-grade. CS Z-Score 120-bar improved from 1.13 to 1.25 Sharpe over two years. TR-MA 120-bar follows at 0.85. The 5-minute timeframe continues to be the most rewarding frequency for systematic strategies.

30-Minute — Mean Reversion Enters the Chat

MR 120-bar now leads at 0.30 Sharpe, dethroning CS Z-Score 120-bar. Two months of sustained MR outperformance have reshaped the full-period statistics. The 30-minute timeframe used to be CS momentum's domain — it's now shared territory.

6-Month Rolling History

How has each strategy performed month-by-month? Consistency matters as much as magnitude.

Daily

5-Minute

30-Minute

What to Watch in March

Three things we're tracking:

Is the MR regime on 30-minute data sustainable? Two months doesn't make a regime, but three would be noteworthy. March's data will be the deciding vote.
Can MR and CS be combined on 30-minute data? Their correlation structure looks favorable — MR thrived in February precisely when CS struggled, and vice versa in earlier months.
The 5-minute momentum machine keeps compounding. CS Z-Score 120-bar improved from 1.13 to 1.25 Sharpe. At what point does this become the consensus trade — and does the edge survive consensus?

Data period: February 2024 - February 2026. Universe: BTC, ETH, SOL, ADA, DOGE, XRP, LTC, DOT, LINK, AVAX, TRX, and others. All strategies are long/short.

Disclaimer: This analysis is for educational and research purposes only. It does not constitute investment advice. Past performance is not indicative of future results. These strategies do not include transaction costs, slippage, or market impact. Always do your own research.

Crypto - Price Action Alpha Report - Jan 2026 — Crypto strategy performance, January 2026
Crypto Strategy League Tables: March 2026 — Which crypto strategies won in March 2026
Crypto Orderflow Alpha Report — Feb 2026 — Order book signals and microstructure, February 2026
Reference Guides - Regime Detection — HMM, change-point detection, and regime-aware trading

Crypto Orderflow Alpha Report — Feb 2026

oracle — Tue, 03 Mar 2026 12:54:22 GMT

Crypto Orderflow Alpha Report — Feb 2026

A systematic evaluation of microstructure-based trading signals on crypto assets.

What This Is

This is not a backtest. There are no strategy equity curves, no hypothetical P&L, and no Sharpe ratios built on frictionless assumptions.

Instead, this is an Information Coefficient (IC) analysis — the same approach used by quantitative hedge funds to evaluate whether a signal contains genuine predictive information before committing capital.

For each signal, we measure its correlation with future price changes across multiple timeframes. A positive IC means the signal has directional predictive power. A high t-statistic means that predictive power is statistically significant, not a fluke of the sample.

This month covers 1–28 February 2026 — 1-second order book snapshots across 6 crypto markets.

The Signals

We extract 13 signals from Level 2 order book snapshots at 1-second frequency. Each captures a different dimension of microstructure information.

Imbalance signals — measure the balance of resting liquidity:

Queue Imbalance Top1 — bid vs ask size at the best price
Queue Imbalance Top5 — bid vs ask depth across the top 5 levels
Depth Imbalance Top10 — bid vs ask depth across the top 10 levels

Derived microstructure signals — higher-order features from the book:

Microprice Bias — volume-weighted mid price deviation from the arithmetic mid
Queue Momentum — rate of change of the top-of-book imbalance
Depth Gradient — how imbalance changes between top 5 and top 10 levels
Slope Imbalance — asymmetry of the book's price-depth slope

Order flow signals — measure the dynamics of order arrivals:

OFI (Order Flow Imbalance) — net order flow at the best bid and ask
OFI Rolling 5s — smoothed OFI over a 5-second window

Composite and conditional signals — combinations and filters:

Combined Imbalance — z-score average of queue imbalance and OFI
Contrarian Imbalance — fades extreme queue imbalances (mean reversion)
Adverse Selection — trades based on adverse selection pressure
Spread Timing — trades queue imbalance only when the spread is tight

Data

1–28 February 2026, 1-second order book snapshots across 6 markets:

BTCUSDT Perpetual + Spot (Binance)
ETHUSDT Perpetual + Spot (Binance)
SOLUSDT Perpetual (Binance)
BNBUSDT Perpetual (Binance)

Forward returns computed at six horizons: 10s, 30s, 60s, 120s, 300s, 600s in basis points.

The Hedge Fund Approach to Alpha Analysis

Before a quant fund trades a signal, it runs it through a battery of statistical tests. We apply the same framework:

Information Coefficient (IC) — Pearson correlation between the signal and the subsequent forward return. Measures linear predictive power.

Rank IC — Spearman rank correlation. More robust to outliers and non-linear relationships.

IC t-statistic — statistical significance of the IC across assets. Values above 2.0 suggest the signal is reliably informative, not a sample artifact.

Hit Rate — fraction of observations where the signal direction matches the return direction. Even a 51% hit rate can be enormously profitable at high frequency.

IC Decay Profile — how predictive power changes across forward horizons. Fast decay means the signal is only useful at very short holding periods.

Signal Autocorrelation — how persistent the signal is over time. High autocorrelation means the signal changes slowly (lower turnover); low means it flips rapidly (higher turnover, higher transaction costs).

Signal Correlation Matrix — which signals are redundant and which are independent. Redundant signals don't add diversification value.
Subscribe now

1. Information Coefficient

The IC table below shows each signal's predictive power at the 60-second forward return horizon, aggregated across all 6 markets.

Queue Imbalance Top1 leads with an IC of 0.043 and a t-statistic of 11.3 — comfortably above the 2.0 threshold for statistical significance.

At the other end, Contrarian Imbalance has a negative IC of -0.074 — meaning this contrarian signal actually moves against the subsequent price direction.

11 of 13 signals have a statistically significant IC (|t-stat| > 2.0). Hit rates cluster in the 49–52% range — small edges, but at 86,400 observations per day per asset, even a 51% hit rate is highly significant.

The key takeaway: order book imbalance signals contain genuine short-term predictive information. The effect is small in absolute terms but statistically robust.

2. IC Decay Profile

How quickly does predictive power decay as we look further forward?

The pattern is clear: imbalance signals peak at the shortest horizons and decay monotonically. This is the signature of genuine microstructure alpha — the information is fleeting and gets priced in quickly.

Queue Imbalance and Microprice Bias signals show the steepest decay — their edge is concentrated in the first 10–60 seconds. OFI-based signals decay more slowly, suggesting they capture a slightly different (and longer-lived) dimension of order flow information.

The Contrarian Imbalance signal is consistently negative across all horizons, confirming that mean reversion of the order book does not predict mean reversion of the price at these timescales.

3. Signal Autocorrelation and Turnover

How persistent are these signals? A signal that flips every second generates enormous transaction costs; one that holds for minutes is cheaper to trade.

Three tiers emerge:

Most persistent — Depth Imbalance Top10, Queue Imbalance Top5, Slope Imbalance. These have high 1-second autocorrelation (>0.95) and low turnover. The deep book changes slowly.

Moderately persistent — Queue Imbalance Top1, Microprice Bias, Adverse Selection. Still highly autocorrelated but with slightly higher turnover as the top of book is more dynamic.

Least persistent — OFI, OFI Rolling 5s, Queue Momentum. These are flow-based signals that fluctuate rapidly. OFI in particular has high turnover, making it expensive to trade naively.

Spread Timing has near-zero turnover because it only trades when the spread is unusually tight — a rare condition.

4. Signal Correlation Matrix

Which signals are telling us the same thing, and which offer independent information?

Several clusters emerge:

Imbalance block — Queue Imbalance Top1, Top5, Depth Imbalance Top10, Microprice Bias, and Adverse Selection are all highly correlated (0.7–1.0). They all measure essentially the same thing: the balance of resting limit orders. Adverse Selection has near-perfect correlation with Queue Imbalance Top1 — they are functionally identical signals.

Order flow cluster — OFI and OFI Rolling 5s correlate with each other but have low correlation with the imbalance block. They capture a different source of information: changes in the book rather than its state.

Independent signals — Queue Momentum, Depth Gradient, and Contrarian Imbalance have low correlation with both clusters. These offer genuine diversification if combined in a multi-signal model.

The practical implication: you only need 3–4 signals to capture most of the available information. One imbalance measure, one OFI measure, queue momentum, and optionally a spread filter.

5. Latency Sensitivity

The final test: how fast do you need to be to capture this alpha?

We evaluate each signal's Sharpe ratio at execution latencies from 0 to 30 seconds. The results are striking:

Queue Momentum is the most latency-sensitive: Sharpe drops from 92 at 0s to 10 at 10s. Its half-life is roughly 2 seconds.

Imbalance signals (Top1, Top5, Microprice Bias) lose about half their value by 5 seconds. At 30s latency, they're barely above zero.

OFI signals are more resilient — OFI Rolling 5s retains meaningful Sharpe even at 10–30s latency. This makes sense: smoothed flow signals change more slowly.

Spread Timing is the only strategy that's essentially latency-insensitive, but it also has the weakest pre-cost Sharpe.

The message is clear: this alpha is real, but it lives in the sub-10-second domain. To exploit it, you need co-located infrastructure with sub-second execution — the domain of professional HFT firms, not retail traders.

Subscribe now

6. Cross-Asset Consistency

Do signals work uniformly or only on specific assets?

The strongest signals (Queue Imbalance, Microprice Bias) show positive IC across all 6 markets. This cross-asset consistency is important — it reduces the risk that we're overfitting to a single instrument's idiosyncrasies.

BTC and ETH perpetual futures tend to show slightly stronger IC than spot markets, consistent with the idea that derivative markets are where informed flow concentrates.

What to Take Away

The alpha is real. Order book imbalance signals have statistically significant predictive power for short-term crypto price movements. The IC values are modest (0.01–0.05) but the t-statistics are strong (3–13), and the results are consistent across assets and time.

But most of it is redundant. Of the 13 signals we tested, you can capture the vast majority of predictable variation with just 3–4: one imbalance measure (Queue Imbalance Top1 or Microprice Bias), one flow measure (OFI), Queue Momentum, and optionally a spread filter.

The alpha decays fast. Most signals lose 50% of their predictive power within 5 seconds of latency. At 30 seconds, nearly all of it is gone. This is not alpha you can capture by placing orders on a web interface.

Mean reversion doesn't work. The Contrarian Imbalance signal — fading extreme book imbalances — has consistently negative IC. When the book is lopsided, the price moves with the imbalance, not against it.

The signals are features, not strategies. No single binary signal here is tradeable after realistic costs (5+ bps round trip). But as inputs to a multi-factor model with smarter entry/exit logic, position sizing, and execution optimization, they contain genuinely useful information.

This is how quant firms use microstructure data: not as standalone trading rules, but as predictive features in a machine learning pipeline that combines multiple weak signals into one stronger one.

Methodology Notes

Forward returns: log(mid_{{t+h}}) − log(mid_t), expressed in basis points
IC: Pearson correlation between signal value and forward return
Rank IC: Spearman rank correlation
t-statistic: IC / (std(IC across assets) / sqrt(N_assets))
All signals computed from Level 2 order book snapshots, resampled to 1-second bars
No look-ahead bias: signals use only data available at time t; forward returns are computed from t+1 onward

For educational purposes only. Not investment advice. Past predictive relationships do not guarantee future results.

5 HFT Secrets Every Quant Trader Should Know — Practical insights from high-frequency trading research
Crypto - Price Action Alpha Report - Jan 2026 — Data-driven analysis of cryptocurrency price patterns
Building a Market-Maker on Hyperliquid — Part III: The Backtester — Building and backtesting a crypto market-making engine

Monthly Crypto Alpha Hunt: Millions of Random Guesses, One Good Trade (Feb 2024-Jan 2026)

oracle — Sun, 08 Feb 2026 06:40:06 GMT

The idea is simple: every month I generate a huge number of random mathematical expressions, evaluate each one as a trading signal on historical crypto data, and see what sticks. No optimization, no gradient descent — pure random search. This month’s standout is a formula on 5M bars that delivered a 2.76 Sharpe over two years. The full report covers Daily, 30M, and 5M results with equity curves.

All alphas are evaluated on a 2-year rolling window (Feb 2024 – Jan 2026) using some of the most liquid crypto assets.

Methodology: Each alpha provides the entry signal. Positions are sized by prediction magnitude and held on a rolling 5-bar window — at any point, the portfolio position reflects the sum of the last 5 predictions. Returns are realized continuously across this window.

Executive Summary

Daily Timeframe

Risk-Adjusted Equity Curves

Performance comparison with volatility normalized to 15% annualized:

30M Timeframe

Risk-Adjusted Equity Curves

Performance comparison with volatility normalized to 15% annualized:

5M Timeframe

Risk-Adjusted Equity Curves

Performance comparison with volatility normalized to 15% annualized:

Monthly Alpha Hunt: Millions of Random Guesses, One Good Trade — Millions of random expressions, one good trade
Monthly Alpha Hunt: March 2026 — Monthly scan for actionable quantitative signals
Simple Alpha Atlas: What Actually Works Across Markets (Daily) — What alpha actually works across markets on daily bars

Crypto - Orderflow Alpha Report - Jan 2026

oracle — Wed, 04 Feb 2026 02:52:40 GMT

What This Is

This is not a backtest. There are no strategies, no hypothetical P&L curves. Instead, we measure something more fundamental: how much information does the order book contain about future price movements?

We compute the Information Coefficient (IC) -- the correlation between a signal observed now and the forward price change in basis points -- across multiple horizons and assets. This is the standard approach used by systematic funds to evaluate alphas before they ever enter a portfolio construction or execution pipeline.

For readers interested in price action alphas, see Price Action Alpha Report - Jan 2025.

The Signals

We extract 13 signals from Level 2 order book snapshots at 1-second frequency:

Imbalance signals (how much more is on the bid vs the ask):

Queue Imbalance Top1: (bid_size - ask_size) / (bid_size + ask_size) at best level
Queue Imbalance Top5: Same, summed over top 5 levels
Depth Imbalance Top10: Same, summed over top 10 levels

Derived microstructure signals:

Microprice Bias: (microprice - mid) in bps, where microprice is the size-weighted fair value
Queue Momentum: 5-second change in queue imbalance
Depth Gradient: Difference between top-5 and top-10 imbalance (near vs far book)
Slope Imbalance: Book shape asymmetry via VWAP distance from mid

Order flow signals:

OFI: Order Flow Imbalance, normalized by total depth
OFI Rolling 5s: 5-second rolling sum of OFI

Composite and conditional signals:

Combined Imbalance: Z-score average of queue imbalance + OFI
Contrarian Imbalance: Signal that fades extreme imbalances (mean reversion)
Adverse Selection: Imbalance weighted by depth at top of book
Spread Timing: Imbalance conditional on tight spread

Data

January 2026, 1-second order book snapshots across 6 markets:

BTCUSDT Perpetual + Spot (Binance)
ETHUSDT Perpetual + Spot (Binance)
SOLUSDT Perpetual (Binance)
BNBUSDT Perpetual (Binance)

Forward returns are computed as log-returns in basis points at horizons of 10s, 30s, 60s, 120s, 300s, and 600s.

The Hedge Fund Approach to Alpha Analysis

Before building a strategy, systematic funds evaluate alphas on several dimensions:

Information Coefficient (IC): Pearson correlation between signal and forward return. An IC of 0.02 is considered meaningful at high frequency.
Rank IC: Spearman rank correlation. More robust to outliers than Pearson IC.
IC t-statistic: Statistical significance. |t| > 2 means the IC is unlikely to be noise.
Hit Rate: Fraction of observations where signal direction matches return direction. >50% means the signal has directional content.
IC Decay Profile: How IC changes across forward horizons. Fast decay = short-lived alpha. Slow decay = more robust signal.
Signal Autocorrelation: How persistent the signal is. High autocorrelation = low turnover = lower execution costs.
Signal Correlation Matrix: Which signals carry independent information vs. being redundant.

1. Information Coefficient

The IC table shows each signal’s predictive power for the 60-second forward return in bps, averaged across all 6 markets.

The imbalance family dominates. Queue Imbalance Top1 has an IC of 0.065 with a t-stat of 10.24 -- highly significant. Microprice Bias is nearly identical (IC 0.061, t-stat 12.17). These signals all measure the same underlying phenomenon: when there is more resting size on the bid, prices tend to rise.

Queue Momentum (IC 0.037) carries less information per observation but has the highest t-stat (13.95), suggesting it is the most statistically robust despite lower IC. This is because it captures the change in the book state rather than the level.

OFI signals have lower IC (0.016-0.023) but are still significant (t > 8). These measure actual order arrivals and cancellations rather than standing depth.

Contrarian Imbalance has IC of -0.117 -- strongly negative. This confirms that in crypto microstructure, queue imbalance is a momentum signal. Fading it loses money consistently, with a hit rate of only 37.6%.

An interesting detail: hit rates for the positive-IC signals are only 51-52%. The signal is barely above coin-flip directionally, but the magnitude of correct predictions outweighs the magnitude of wrong ones. This is typical of microstructure alphas -- small edge, applied many times.

2. IC Decay Profile

The IC decay curve is the most important diagnostic. It tells you how quickly information gets priced in.

Queue Imbalance Top1 peaks at IC 0.129 at the 10-second horizon, decays to 0.065 at 60s, and still holds 0.023 at 600s (10 minutes). The decay is smooth and monotonic -- classic microstructure alpha behavior.

All imbalance signals follow a similar pattern: peak IC at 10s around 0.12, halving roughly every 60 seconds. By 600s the IC is about one-fifth of its peak, but still positive and significant.

Queue Momentum peaks at 0.081 (10s) and decays faster -- its information is more transient, consistent with it measuring a derivative (rate of change) rather than a level.

Spread Timing behaves differently from all other signals. Its IC actually increases from 10s (0.080) to 120s (0.104), suggesting it captures a slower-moving phenomenon. This makes sense: tight spreads tend to persist and the imbalance signal conditional on tight spreads has a longer shelf life.

Contrarian Imbalance is consistently negative across all horizons (IC -0.227 at 10s, -0.036 at 600s). Mean reversion in the order book does not work at any timescale we tested.

3. Signal Autocorrelation and Turnover

Signal persistence matters enormously for implementation. A signal that changes every second requires thousands of trades per day; one that persists for minutes requires far fewer.

The autocorrelation table reveals a clear hierarchy:

Most persistent: Depth Imbalance Top10 (AC at 1s: 0.75, at 30s: 0.14) -- deeper book levels change slowly
Moderately persistent: Queue Imbalance Top1 (AC at 1s: 0.72, at 30s: 0.08) -- best level turns over faster
Least persistent: OFI Normalized (AC at 1s: -0.02) -- effectively white noise at 1-second frequency, turnover of 54%

Queue Momentum has negative autocorrelation at 5s (-0.20), meaning the signal tends to reverse itself over 5 seconds. This explains its high turnover (26%) and is consistent with the derivative of a mean-reverting process.

Spread Timing has the lowest turnover (0.07%) -- it barely changes. This is because it is zero most of the time and only activates when spreads are unusually tight.

4. Signal Correlation Matrix

The correlation matrix answers the key question: which signals carry independent information?

Three distinct clusters emerge:

Cluster 1: The imbalance block (correlations 0.88-1.00)
Queue Imbalance Top1, Top5, Depth Imbalance Top10, Microprice Bias, Combined Imbalance, and Adverse Selection are all measuring essentially the same thing. Adverse Selection has a perfect 1.00 correlation with Queue Imbalance Top1 -- it adds zero independent information. Including multiple signals from this cluster in a model is redundant.

Cluster 2: Order flow (correlations 0.17-0.27 with Cluster 1)
OFI Normalized and OFI Rolling 5s correlate only 0.17 with imbalance. This is meaningful -- order flow captures information about changes in the book that standing depth does not. A model using both imbalance and OFI should have higher combined IC than either alone.

Cluster 3: Independents
Queue Momentum (0.30 with imbalance), Depth Gradient (0.18), and Spread Timing (0.02) are the most independent signals. Spread Timing in particular is essentially uncorrelated with everything -- it captures a different phenomenon entirely.

Slope Imbalance is mildly negatively correlated (-0.11) with imbalance, and it has negative IC. This confirms it is noise in this market structure.

5. Latency Sensitivity

This is the money chart. It shows what happens to your signal as execution latency increases from 0 to 30 seconds.

The half-life column tells the story: most imbalance signals lose 50% of their predictive power within 5 seconds. Queue Momentum decays fastest (half-life 2s). Spread Timing is the most latency-robust (half-life 30s).

At 30 seconds of latency, the imbalance signals still retain some predictive content, but it’s about 8-10% of the original level. For a systematic fund with co-located infrastructure, the 0-2s column is the relevant one. For anyone else, the 5-30s columns show what’s realistically accessible.

What to Take Away

The order book contains real, statistically significant information about future price movements in basis points. The strongest signals are simple: queue imbalance at the top of book, with IC around 0.065 at 60 seconds and 0.13 at 10 seconds.

But the alpha research framework reveals the nuances:

Most of these signals are highly redundant -- you need at most 3-4, not 13
The useful independent dimensions are: imbalance level (pick one), OFI, queue momentum, and spread regime
IC decays roughly 50% every 60 seconds, reaching near-zero by 10 minutes
Mean reversion has strongly negative IC -- imbalance is a pure momentum signal
Deeper book levels (Top10) are more persistent but have lower IC than Top1

The practical takeaway: these signals are most valuable as features in a broader model -- adjusting fair value estimates, timing execution, or adding a short-term momentum component to longer-horizon predictions.

Methodology Notes

Forward returns: log(mid_{t+h}) - log(mid_t), in basis points (x 10,000)
IC: Pearson correlation between signal and forward return, averaged across assets
Rank IC: Spearman rank correlation
t-stat: IC / (std(IC across assets) / sqrt(N_assets))
All signals computed from L2 snapshots resampled to 1s
No look-ahead: signals use data at time t, returns measured from t+1

Disclaimer: For educational purposes only. Not investment advice.

Crypto Orderflow Alpha Report — Feb 2026 — Order book signals and microstructure, February 2026
HFT Secrets 3/5: Microprice — The Fair Value Hidden in the Order Book — Volume-weighted fair value from the order book
HFT Secrets 1/5: Order Flow Imbalance — Reading the Tape in Real Time — Computing and trading order flow imbalance
Reference Guides - Order Book Dynamics — Queue dynamics, price formation, and microstructure

Crypto - Price Action Alpha Report - Jan 2026

oracle — Mon, 02 Feb 2026 05:36:17 GMT

The goal is simple: find what actually works in crypto — and what doesn’t. Let’s see what January had to say.

For readers interested in microstructure alphas, see Orderflow Alpha Report

January’s Winners

Cut through the noise. Here’s what worked best at each timeframe:

Three timeframes, three different winners. That’s not a bug — it’s the market telling you that timeframe selection matters as much as strategy selection.

How to Read This

Each winner tells a different story. Daily rewarded patience — the 120-day moving average captured the macro trend. 30-minute rewarded contrarians — mean reversion worked because intraday moves overshot and snapped back. 5-minute rewarded speed — fast breakouts on the shortest lookback. If you’re only trading one timeframe, you’re leaving edge on the table.Green = positive risk-adjusted returns. Red = negative. Use this table to spot regime shifts and timeframe divergences at a glance.

The 4 Strategy Types

TR-BO (Breakout) — Long above N-bar high, short below N-bar low. The classic channel breakout.
TR-MA (Moving Average) — Long above the N-bar moving average, short below. Simple, old-school, effective.
CS (Cross-Sectional Momentum) — Rank all 19 coins by relative strength. Buy the leaders, sell the laggards.
MR (Mean Reversion) — Z-Score based. Buy the oversold, short the overbought. The contrarian’s bet.

Lookback windows: 5, 20, and 120 bars. Universe: 19 major cryptocurrencies. All strategies are long/short.

January 2026: The Scoreboard

Daily Strategies

Don't let the final Sharpe fool you — January was choppy and inconclusive for most of the month. The equity curves tell the real story: strategies oscillated around zero for three weeks before trend-following finally pulled ahead in the last week. If you'd given up mid-month, you'd have missed the payoff.

But look at the bottom of the table. Mean reversion got obliterated. MR 120-day posted -5.89 Sharpe, MR 20-day at -2.29, and MR 5-day at -1.16. When crypto trends, the contrarians pay dearly.

The takeaway: January’s crypto market had strong directional moves. Trend strategies captured them on both sides (long and short). Mean reversion strategies fought the tape and lost.

30-Minute Strategies

Now here’s where it gets interesting. On 30-minute bars, the entire playbook flips.

Mean reversion swept the podium: MR 20-bar at +0.55, MR 120-bar at +0.51, MR 5-bar at +0.45. Meanwhile, every single trend strategy posted negative Sharpes, with TR-BO 20-bar the worst at -0.71.

The paradox: The same market, the same month, the same coins — but zoom into 30-minute bars and the winners become losers. This is why timeframe diversification isn’t optional. It’s essential.

5-Minute Strategies

The 5-minute timeframe tells a quieter story. Returns are compressed (this is microstructure territory), but the signal is real.

TR-BO 5-bar led at +2.61 Sharpe, followed by CS Z-Score 20-bar at +2.41 and CS 120-bar at +2.08. Trend strategies dominated the top half of the table. Mean reversion was the only loser, with all MR strategies posting negative Sharpes.

The Timeframe Paradox

Here’s the part that should make you rethink everything you know about “what works in crypto”:

On daily bars, TR-MA posted a 3.63 Sharpe. On 30-minute bars, TR-MA posted -0.55 Sharpe. Mean reversion posted -5.89 Sharpe on daily bars but +0.55 Sharpe on 30-minute bars.

This isn’t noise. This is the fundamental structure of the market operating at different frequencies. Daily bars capture macro regime shifts — rate decisions, regulatory news, ETF flows. 30-minute bars capture intraday mean reversion — the market’s tendency to overshoot and snap back within the same session.

Neither timeframe is “right.” Both are real. The edge is in knowing when each applies.

Performance by Asset

Which coins rewarded which strategies? The asset-level tables show the best-performing horizon for each strategy type per coin.

January 2026 (Daily)

EOS was the cross-sectional momentum darling, posting +9.1% on CS alone. DOT and ATOM were the trend-followers’ best friends, with broad green across TR-BO and TR-MA. MATIC and SHIB are flatlined at 0.0% across the board — essentially dead coins from a strategy perspective.

6-Month Total (Daily)

The 2-Year View: Who Survives?

Monthly returns are exciting. But the real question is: which strategies hold up over two years? These are risk-adjusted returns (Sharpe ratios) since January 2024.

Daily — Trend Rules the Long Game

TR-BO 120-day leads the daily leaderboard at 0.63 Sharpe. All five top-ranked strategies are either trend or momentum. Not a single mean reversion strategy makes the cut. Over two years, on daily bars, fighting the trend is a losing proposition.

5-Minute — Momentum’s Kingdom

The single best risk-adjusted strategy across all timeframes. CS Z-Score 120-bar on 5-minute data has maintained a 1.13 Sharpe over two years. That’s institutional-grade. TR-MA 120-bar follows at 1.03 — a result that was invisible when trend strategies were long-only. The long/short upgrade unlocked hidden alpha on this timeframe.

30-Minute — The Comeback Kid

TR-BO 120-bar leads at 0.84 Sharpe — making it the second-best strategy across all timeframes. The 30-minute timeframe used to be a wasteland for trend-following when strategies were long-only. The addition of the short side transformed it into one of the most attractive frequencies.

What to Watch in February

Three things we’re tracking:

Will the daily trend regime persist? January’s 3.63 Sharpe on TR-MA 120-day is exceptional. If macro conditions shift, mean reversion could snap back hard.
The 5-minute momentum machine. CS Z-Score 120-bar has been quietly compounding at a 1.13 Sharpe for two years. Can anything dethrone it?
MATIC and SHIB. Two coins showing zero signal across all strategies. Are they simply range-bound, or is liquidity drying up? Worth watching.

Monthly Alpha Hunt: Millions of Random Guesses, One Good Trade — Monthly scan for actionable quantitative trading signals
How my hyperliquid vault trades: Multi-Horizon Systematic Trading — How our Hyperliquid vault combines multi-timeframe systematic signals
Crypto Orderflow Alpha Report — Feb 2026 — Decoding crypto market microstructure through order flow

Delphic Alpha: Alpha Reports

Alpha Combinations, Part 2: From Theory to Practice

Which Pairs Create the Most Value?

Top Pairs: Equity Curves

From Pairs to Full Portfolios

L1 vs L2: A Visual Intuition

What the Weights Look Like

The IS vs OOS Test

Tuning the Penalty

Equity Curves

Going Cross-Asset

Measuring the Diversification Benefit

Takeaways

The Exotic Alpha Atlas: Beyond Textbook Signals on 30-Minute Bars

Signal Definitions

The Exotic Alpha × Asset Class Heatmap

Best Exotic Alpha per Asset Class

Exotic vs Textbook: Does Complexity Pay?

The Overfitting Tax: Exotic Edition

Equities: 19% Shrinkage (slope 0.81)

FX: 25% Shrinkage (slope 0.75)

Bonds: 43% Shrinkage (slope 0.57)

Crypto: 64% Shrinkage (slope 0.36)

Commodities: −163% Shrinkage (slope −0.63)

Regime Dependence

Cross-Asset Diversification

Equity Curves by Asset Class

Equity Index Futures

Bond Futures

Commodity Futures

G7 FX

Crypto

Takeaways

Related Articles

The Intraday Alpha Atlas: What Works on 30-Minute Bars

Signal Definitions

The Alpha × Asset Class Heatmap

Best Strategy per Asset Class

The Overfitting Tax

The Tradability Lens: Turnover, Holding Period, and Edge per Trade

Regime Dependence

Diversification at Higher Frequency

Equity Curves by Asset Class

Equity Index Futures

Bond Futures

Commodity Futures

G7 FX

Crypto

Takeaways

Related Articles

Monthly Alpha Hunt: March 2026

Executive Summary

Daily Timeframe

Risk-Adjusted Equity Curves

30M Timeframe

Risk-Adjusted Equity Curves

5M Timeframe

Risk-Adjusted Equity Curves

Recommendation

How to Read the Genes

This Month's Winners — In Plain English

Related Articles

Crypto Strategy League Tables: March 2026

March’s Winners

The 4 Strategy Types

Daily Strategies

30-Minute Strategies

5-Minute Strategies

The Timeframe Paradox

Performance by Asset

Daily

30-Minute

5-Minute

The 2-Year View: Who Survives?

Daily — Trend Still Rules, but Cracks Appeared

5-Minute — Momentum’s Kingdom

30-Minute — Mean Reversion Consolidates

6-Month Rolling History

What to Watch in April

Related Articles