Subscribe now

This is infrastructure, not a black box. The repo gives you the plumbing: data pipelines, execution, risk management, monitoring. The alpha comes from you. The strategy templates are starting points with promising backtests, but what you build on top of them is what matters.

Not investment advice. Past backtest performance does not guarantee future results. You are solely responsible for your trading decisions and any losses incurred.

What is the Delphic Trading Desk?

A private GitHub repository for annual subscribers. Mid-frequency systematic trading infrastructure. strategy templates, alpha datasets, and live trading code. Backtest strategies, deploy them live, and monitor everything.

This is not a library you import. It is a complete trading system: data collection, signal generation, order execution, risk management, and operations monitoring. You clone it, run a backtest, and deploy to production with the same code.

The first connector is Hyperliquid. crypto perpetual futures, 200+ coins, no KYC for non-US users. MT5 and cTrader connectors are coming next. These open the door to forex, indices, and commodities. and critically, to prop firm challenges and platforms like Darwinex where you can build a verified track record and attract investor capital. Same signal interface, same risk engine. write once, deploy anywhere.

The repo ships with three strategy templates to start with. mean reversion, buy the dip, and a feature alpha scanner. each with promising backtest characteristics. More are coming, and subscribers can contribute their own via pull request.

What's Inside

Data pipeline. Three ways to get Hyperliquid data: REST API (no auth, paginates automatically), S3 tick archive (millisecond fills, needs AWS), and a real-time WebSocket streamer that aggregates trades into OHLCV bars with automatic gap backfill on reconnect. 76 liquid coins tracked by default.

Feature engine. 68 feature families. 336 features from OHLCV data. RSI, MACD, ATR, Bollinger Bands, z-scores, candle shapes, breakout signals, cross-sectional ranks, and 9 atlas signal families from Delphic Alpha research. Same engine used in production. Works on any timeframe, any asset class.

Three strategy templates. Each downloads data, runs a backtest, prints results, and shows the exact commands to go live:

• Mean reversion. z-score fade. When a coin moves 3+ standard deviations from its 4-day moving average, trade the other way. Sharpe 5.28 gross on 1H bars across 76 coins.

• Buy the dip. buy when price drops 12% from its 7-day high, hold 4 hours. Vol-scaled entries.

• Feature alpha scanner. scan all 336 features against next-bar returns, rank by Sharpe, build a composite signal from the best ones.

Execution. Two modes: market orders (immediate) or passive-first (post limit at bid/ask, wait 10 seconds, cross if unfilled. saves ~7 bps per round trip). Handles position sync, flip logic, coin-specific size rounding, and graceful shutdown.

Risk engine. Pluggable checks that run before every trade. Built-in: max exposure, max position per coin, daily loss limit, drawdown kill switch, concentration limit. Write your own in Python. Or use CLI flags:

python trader.py --signal my_signal.py \
    --max-exposure 500 --max-position 100 --max-daily-loss 50

Operations. Every signal and fill is logged to CSV. Position snapshots every 60 seconds. The trade monitor tracks account value, unrealized PnL, margin, and drawdown from peak. If drawdown exceeds your threshold, it flattens everything and writes a kill switch file that stops the trader from re-opening positions.

Supervisors. Bash scripts that keep everything alive. Streamer crashed? Restarted in 30 seconds. Trader hung? Detected by data freshness, killed, restarted. Each run gets its own timestamped log directory. PID files prevent accidental duplicates. One streamer feeds multiple traders.

From Backtest to Live in 4 Commands

# 1. Set up credentials
python connectors/hyperliquid/setup_credentials.py

# 2. Backtest
python connectors/hyperliquid/strategies/mean_reversion_template.py \
    --coins BTC,ETH,SOL --days 60

# 3. Dry-run (full loop, no real orders)
python connectors/hyperliquid/trader.py \
    --signal connectors/hyperliquid/strategies/mean_reversion_template.py \
    --interval 1h --coins BTC,ETH,SOL --dry-run

# 4. Go live
python connectors/hyperliquid/trader.py \
    --signal connectors/hyperliquid/strategies/mean_reversion_template.py \
    --interval 1h --coins BTC,ETH,SOL \
    --base-bet 50 --max-exposure 500

Production Deployment

For long-running live trading, use the supervised setup. The streamer collects bars via WebSocket, the trader reads from CSV files (instant, no API rate limits), and the supervisor auto-restarts on crash.

# Terminal 1: stream 1H bars for 76 coins, backfill 200 bars of history
INTERVAL=1h COINS_FILE=connectors/hyperliquid/strategies/coins.txt BACKFILL=200 \
    bash connectors/hyperliquid/streamer_supervisor.sh

# Terminal 2: mean reversion, $50 per position, live
SIGNAL=connectors/hyperliquid/strategies/mean_reversion_template.py \
    INTERVAL=1h COINS_FILE=connectors/hyperliquid/strategies/coins.txt \
    MODE=live DATA_SOURCE=csv BASEBET=50 \
    bash connectors/hyperliquid/trader_supervisor.sh

# Terminal 3: buy the dip (same streamer, second strategy)
SIGNAL=connectors/hyperliquid/strategies/buy_the_dip_template.py \
    INTERVAL=1h COINS_FILE=connectors/hyperliquid/strategies/coins.txt \
    MODE=live DATA_SOURCE=csv BASEBET=50 \
    bash connectors/hyperliquid/trader_supervisor.sh

Write Your Own Strategy

Every strategy is one function. The trader handles everything else. data loading, execution, risk checks, journaling.

NOTIONAL = 50  # USD per position

def signal(data: dict) -> dict:
    """
    data   = {coin: DataFrame with open/high/low/close/volume}
    return = {coin: target_notional_usd}
    """
    targets = {}
    for coin, df in data.items():
        if len(df) < 30:
            continue
        close = df["close"]
        z = (close.iloc[-1] - close.rolling(20).mean().iloc[-1]) / close.rolling(20).std().iloc[-1]
        if z < -2.0:
            targets[coin] = NOTIONAL    # long $50
        elif z > 2.0:
            targets[coin] = -NOTIONAL   # short $50
    return targets

Save it as my_signal.py, test it, go live:

python trader.py --signal my_signal.py --interval 1h --coins BTC,ETH,SOL --test
python trader.py --signal my_signal.py --interval 1h --coins BTC,ETH,SOL --dry-run
python trader.py --signal my_signal.py --interval 1h --coins BTC,ETH,SOL

Claude Code Integration

The repo includes SKILL.md files in every folder. Open it in Claude Code and ask it to write a strategy. it already knows the signal interface, size rounding rules, minimum order limits, risk engine API, and the full deployment pipeline. You spend time on alpha research, not plumbing.

Access

The Delphic Trading Desk is available to annual subscribers. Annual plans are 25% off until Sunday. that's the best deal we'll offer on this.

The three companion posts walk through each layer:

• The Hyperliquid Trading Connector. architecture, execution, risk engine

• Building Alpha Datasets: yfinance and Hyperliquid. data pipeline, feature engine, alpha scan

• Three Strategy Templates. mean reversion, buy the dip, feature alpha scanner with backtest results

If you're already an annual subscriber, reply to this post with your GitHub username and I'll add you.

The feature engine takes any OHLCV DataFrame — equities from yfinance, crypto from Hyperliquid, futures, FX — and computes 336 features across 60 families. Same code, same API, different data. Two worked examples below.

Example 1: Daily Equities from yfinance

Download SPY, compute features, scan for alpha.

import yfinance as yf
from feature_engine.features.base import FeatureEngine
from feature_engine.features.registry import get_feature_registry
from feature_engine.config.dataset import FeatureConfig

# Download
df = yf.download("SPY", period="2y", auto_adjust=True)

# Compute all 336 features
families = sorted(get_feature_registry().keys())
config = FeatureConfig(
    families=families,
    periods=[5, 10, 20],
    price_col="Close",
)
engine = FeatureEngine(config)
result_df, feature_columns = engine.compute_all(df)

# Add forward return targets
for h in [1, 5, 10, 20]:
    result_df[f"fwd_ret_{h}"] = result_df["Close"].pct_change(h).shift(-h)

Or from the command line:

python feature_engine/example_yfinance_features.py \
    --tickers SPY,QQQ,IWM,GLD,TLT --period 2y

Feature engine: 60 families, periods=[5, 10, 20]

============================================================
  SPY
============================================================
  Bars: 502 (2024-07-30 to 2026-07-30)
  Features: 336
  Targets: ['fwd_ret_1', 'fwd_ret_5', 'fwd_ret_10', 'fwd_ret_20']
  Valid rows: 502 / 502

Top standalone alphas — ranked by |IC| against 5-day forward returns, daily long/short PnL annualized:

Feature                                IC      Sharpe   RPT
-------------------------------------  ------  ------   --------
Bollinger Bandwidth (20)               +0.230  +2.79    +93 bps
Parkinson Vol (5)                      +0.221  +2.29    +82 bps
Garman-Klass Vol (5)                   +0.218  +2.85   +104 bps
Bollinger Bandwidth (5)                +0.215  +3.17   +113 bps
Realized Vol (5)                       +0.213  +3.18   +110 bps

All five are volatility estimators with positive IC — buy after high vol, sell after low vol. This is the well-documented volatility mean reversion effect on daily equities: high-vol regimes precede recoveries. Sharpe ratios of 2-3 on a 2-year single-stock sample are strong but in-sample.

Example 2: 5-Minute Crypto from Hyperliquid

Same feature engine, different data source. The Hyperliquid REST API provides free candle data for all perpetual contracts — no API key, no account needed.

python feature_engine/example_hyperliquid_features.py \
    --coins BTC,ETH --interval 5m --days 30

Downloading 5m candles (30 days)...
  Downloaded 2 coins

Computing features (60 families, periods=[5, 10, 20])...
  BTC: 5023 bars, 336 features
  ETH: 5023 bars, 336 features

============================================================
  DATASET SUMMARY
============================================================
  Coins:      2
  Total rows: 10046
  Features:   336
  Targets:    ['fwd_ret_1', 'fwd_ret_5', 'fwd_ret_10', 'fwd_ret_20']

For deeper history, the S3 tick archive has millisecond-level fills going back years. Requires AWS CLI and lz4:

python connectors/hyperliquid/download_s3.py \
    --coins BTC,ETH,SOL --days 90 --interval 5m --profile hl-data

Top standalone alphas — same scan, 5-bar forward returns on 5m data:

Feature                                IC      Sharpe   RPT
-------------------------------------  ------  ------   --------
Asym Vol Ratio (20)                    -0.081  +1.99    +99 bps
Max Value Swing Low (10)               -0.080  +7.76   +379 bps
Max Value Mid-Low (10)                 -0.079  +2.74   +129 bps
TRIX (10)                              -0.077  +6.22   +296 bps
Aroon Oscillator (20)                  -0.076  +4.85   +186 bps

Different universe, different features. The top crypto alphas are microstructure and momentum signals — asymmetric volatility, swing point structure, TRIX — all with negative IC (buy when indicator is low = mean reversion). No overlap with the equity vol features. The same engine surfaces different edges on different data.

What Works Where

Daily equities: volatility features dominate (Bollinger width, Parkinson, Garman-Klass, realized vol). The signal is simple: buy when vol is elevated, markets tend to recover.

5-minute crypto: price action and momentum features dominate (swing structure, TRIX, Aroon). The signal: fade moves after extreme momentum readings.

Both are mean reversion — they just measure different things. The feature engine computes all 336 features regardless; you scan and keep what works on your data.

The Feature Families

60 families, grouped by category:

Momentum:       RSI, MACD, ROC, Stochastic, CMO, CCI, Williams %R, TRIX, Aroon
Trend:          ADX, SAR, Linear Regression, Hilbert Transform
Volatility:     ATR, Bollinger, Parkinson, Garman-Klass, Yang-Zhang, Keltner
Volume:         OBV, MFI, AD, Volume z-scores, Accumulation/Distribution
Price action:   Candle shape, gaps, wick ratios, breakout direction
Statistical:    Z-scores, skewness, rolling quantiles, mean reversion speed
Composite:      Fractal trend, fade breakout, Keltner reversion (Atlas signals)

Each family is computed at every lookback period you specify (default: 5, 10, 20 bars). Families that need intraday resolution or multiple assets (session opens, cross-sectional ranks) are automatically skipped when the data doesn't support them.

The Pipeline

Step 1 is the same for both sources. Steps 2-3 are Hyperliquid-specific (cross-sectional ranking across coins):

# 1. Build features (pick your source)
python feature_engine/example_yfinance_features.py --tickers SPY,QQQ --period 5y
# or
python feature_engine/example_hyperliquid_features.py --coins BTC,ETH --days 60

# Output: features.parquet — ready for any ML model or alpha scan

# 2. Scan for standalone alphas (Hyperliquid — ranks across coins)
python connectors/hyperliquid/strategies/feature_alpha_template.py

# 3. Go live with what works (Hyperliquid)
python connectors/hyperliquid/trader.py \
    --signal connectors/hyperliquid/strategies/feature_alpha_template.py \
    --interval 1h --passive --max-exposure 200 --max-daily-loss 25

The scanner ranks features by IC across all coins, picks the top 5, and saves config.json. The trader loads it automatically. For yfinance data, the parquet is the end product — feed it to your own model or alpha scan.

AI-Assisted Research

Every folder in the repo has a SKILL.md file that teaches AI coding assistants (Claude Code, Cursor, Copilot) how to use the code. Open the repo and ask "download 60 days of BTC 4H data and scan all features for alpha" — the AI knows the commands, the interfaces, which families to skip on daily data, and what the common errors are.

All of this is in the Delphic Trading Desk repo.

Related posts: The Hyperliquid Trading Connector (infrastructure + deployment) and Three Strategy Templates for Hyperliquid (backtesting + going live).

Disclaimer: This is not investment advice. All results are backtested and in-sample. Past performance does not predict future results.

Each template downloads data, runs a backtest, and prints the exact commands to go live. Modify the parameters, swap the signal logic, add your own features. The base is here.

The repo includes coins.txt — a list of 76 Hyperliquid perps with over $500K daily volume. One coin per line, easy to edit. All templates and scripts read from this file via --coins-file coins.txt so you define your trading universe in one place.

All three templates accept the same data flags: --days for the REST API (default), --s3 --profile hl-data for the S3 tick archive, or --from-csv to load pre-downloaded CSVs.

A note on data: the REST API is unauthenticated and free, but rate-limited to roughly 20 requests per second. Each coin needs one request per 5,000 bars, so downloading 76 coins at 6 months of 1H data takes about 5 minutes. The API currently serves around 6 months of 1H history (back to late January 2026). For deeper history or tick-level precision, the S3 archive goes further but downloads the full exchange feed per hour (~40 MB compressed), so it is much slower. All downloads are cached to CSV — the second run loads instantly.

Mean Reversion

Rank all coins by z-score. Long the most oversold, short the most overbought. Rebalance every bar.

Backtest

# REST API (fast, no auth):
python connectors/hyperliquid/strategies/mean_reversion_template.py \
    --coins-file coins.txt --days 180

# S3 tick data (download first, then load):
python connectors/hyperliquid/download_s3.py \
    --coins-file coins.txt --interval 1h --profile hl-data
python connectors/hyperliquid/strategies/mean_reversion_template.py \
    --from-csv data/hyperliquid/

The script downloads 1H candles, runs a backtest, and shows the current signal — which coins are oversold or overbought right now.

The top row is the default: 100-hour lookback, z > 3.0, single most extreme coin per side. Sharpe 5.28 gross, 7.7 bps per trade, 1,184 trades over 6 months. The pattern is clear: longer lookbacks, higher thresholds, and fewer positions per side all improve the signal. These are gross returns — subtract your own execution costs.

Parameters you can tune:

• LOOKBACK — z-score window (default: 100 bars)

• THRESHOLD — minimum |z| to trade (default: 3.0)

• TOP_N — positions per side (default: 1)

• NOTIONAL — USD per position (default: $11, Hyperliquid minimum is $10)

Go Live

# Test — run signal once, print targets, no orders
python connectors/hyperliquid/trader.py \
    --signal connectors/hyperliquid/strategies/mean_reversion_template.py \
    --interval 1h --coins-file coins.txt --test

# Live with risk limits and passive execution
python connectors/hyperliquid/trader.py \
    --signal connectors/hyperliquid/strategies/mean_reversion_template.py \
    --interval 1h --coins-file coins.txt --passive \
    --max-exposure 200 --max-daily-loss 25

# Production (supervisor restarts on crash)
SIGNAL=connectors/hyperliquid/strategies/mean_reversion_template.py \
    INTERVAL=1h COINS=$(cat coins.txt | tr "
" ",") MODE=live \
    nohup bash connectors/hyperliquid/trader_supervisor.sh &

Buy the Dip

Enter when a coin drops 12% from its 7-day high. Hold for 4 hours — catch the sharp bounce, don't marry the position. Vol-scaled sizing.

Backtest

# REST API (default, fast):
python connectors/hyperliquid/strategies/buy_the_dip_template.py \
    --coins-file coins.txt --days 180

# S3 tick data (slower, tick-accurate):
python connectors/hyperliquid/strategies/buy_the_dip_template.py \
    --s3 --profile hl-data --days 30

The parameters below came from a sweep across 76 coins and 6 months of 1H data — the key finding was that short holds (4 hours) after deep dips (-12%) massively outperform longer holds.

The top row is the default: 7-day lookback, 4-hour hold, -12% threshold. Sharpe 2.90 gross, 53% win rate, 3,056 trades. Short holds after deep dips consistently outperform longer holds. The bottom rows show what happens with shallow thresholds and long holds — the edge disappears. All returns are gross.

Parameters:

• LOOKBACK — bars for rolling high (default: 168 = 7 days at 1H)

• HOLD — bars to hold after entry (default: 4 = 4 hours)

• THRESHOLD — dip threshold in % (default: -12%)

• MAX_CONCURRENT — max positions at once (default: 5)

Go Live

# Test
python connectors/hyperliquid/trader.py \
    --signal connectors/hyperliquid/strategies/buy_the_dip_template.py \
    --interval 1h --coins-file coins.txt --test

# Live
python connectors/hyperliquid/trader.py \
    --signal connectors/hyperliquid/strategies/buy_the_dip_template.py \
    --interval 1h --coins-file coins.txt --passive \
    --max-exposure 200 --max-daily-loss 25

# Production
SIGNAL=connectors/hyperliquid/strategies/buy_the_dip_template.py \
    INTERVAL=1h COINS=$(cat coins.txt | tr "
" ",") MODE=live \
    nohup bash connectors/hyperliquid/trader_supervisor.sh &

Feature Alpha Scanner

Scan 336 features as standalone alphas, rank by Sharpe ratio, and build a composite signal from the best ones.

Backtest

# Step 1: build features
python feature_engine/example_hyperliquid_features.py \
    --coins-file coins.txt --days 180

# Step 2: scan and rank
python connectors/hyperliquid/strategies/feature_alpha_template.py

The scanner tests each of the 336 features (68 families at lookbacks 5, 10, 20) against next-bar returns. It outputs a ranked CSV and saves config.json with the best features and their directions.

336 features scanned in seconds. Volatility estimators, Bollinger Bands, and candle shape features show the strongest standalone Sharpe on hourly crypto. Negative Sharpe features are equally useful — they tell you what to fade. The trader loads config.json automatically when you go live.

Go Live

# Test
python connectors/hyperliquid/trader.py \
    --signal connectors/hyperliquid/strategies/feature_alpha_template.py \
    --interval 1h --coins-file coins.txt --test

# Live
python connectors/hyperliquid/trader.py \
    --signal connectors/hyperliquid/strategies/feature_alpha_template.py \
    --interval 1h --coins-file coins.txt --passive \
    --max-exposure 200 --max-daily-loss 25

For 5-Minute and 15-Minute Strategies

The examples above use 1H bars. For faster intervals, you need the WebSocket streamer feeding bars in real-time:

# Start the streamer (separate process)
INTERVAL=15m COINS=BTC,ETH,SOL \
    nohup bash connectors/hyperliquid/streamer_supervisor.sh &

# Trader reads from streamer CSVs
python connectors/hyperliquid/trader.py \
    --signal connectors/hyperliquid/strategies/mean_reversion_template.py \
    --interval 15m --coins-file coins.txt --data-source csv --data-dir data/stream

The streamer backfills gaps automatically on reconnect.

All of this is in the Delphic Trading Desk repo.

Related: The Hyperliquid Trading Connector (infrastructure + deployment).

Disclaimer: This is not investment advice. All results are backtested on historical data. Past performance does not predict future results.

This is the first connector in the Delphic Trading Desk - a private GitHub repo for annual subscribers with mid-frequency trading infrastructure. More connectors (cTrader, Interactive Brokers, more crypto exchanges, user requests) are coming. This one covers Hyperliquid perpetual futures: download data, backtest strategies, and go live.

This connector is the result of multiple iterations and incorporates institutional-grade ideas: passive-first execution with market fallback, pluggable risk checks, automatic gap backfill, process supervision with hang detection, and coin-specific size rounding. It has a risk engine built in - max exposure, max position, daily loss limits, drawdown kill switch - because the one thing worse than a bug is a bug with no position limits.

The repo includes SKILL.md files that encode this domain knowledge. Open it in Claude Code and ask it to write a strategy - it already knows the signal interface, how to handle size rounding and minimum order limits, and the full deployment pipeline. You spend time on alpha research, not plumbing.

Subscribe now

The Signal Interface

Every strategy is one function:

def signal(data: dict) -> dict:
    """
    data  = {coin: DataFrame with open/high/low/close/volume}
    return = {coin: target_notional_usd}
    """
    targets = {}
    for coin, df in data.items():
        close = df["close"]
        z = (close.iloc[-1] - close.rolling(20).mean().iloc[-1]) / close.rolling(20).std().iloc[-1]
        if abs(z) > 2.0:
            targets[coin] = -z * 100  # fade the move
    return targets

The trader loads your function, calls it at every bar close, diffs against current positions, and executes. Three modes: --test (run once, no orders), --dry-run (full loop, no orders), or live.

The Streamer Supervisor

For 5-minute and 15-minute strategies, you need real-time bar data. The streamer supervisor manages this:

INTERVAL=15m COINS=BTC,ETH,SOL \
    nohup bash connectors/hyperliquid/streamer_supervisor.sh &

It runs and monitors one process:

• streamer.py - connects to the Hyperliquid WebSocket, subscribes to trades for all configured coins, aggregates them into OHLCV bars, writes CSV files. File locking for safe concurrent reads. Exponential backoff on disconnection. Stale detection - reconnects if no message for 2 minutes. On reconnect, automatically backfills missing bars from the REST API. A background thread checks for gaps every two bar periods during normal operation.

The supervisor checks every 30 seconds. If the streamer crashes, it restarts it. If it hangs (no log output for 5 minutes), it kills and restarts. If the internet goes down, it pauses and restarts everything on recovery.

For 1H and 4H strategies, you do not need the streamer. The trader fetches directly from the REST API each cycle.

The Trader Supervisor

The trader supervisor manages two processes:

SIGNAL=connectors/hyperliquid/strategies/mean_reversion_template.py \
    INTERVAL=4h COINS=BTC,ETH,SOL MODE=live \
    nohup bash connectors/hyperliquid/trader_supervisor.sh &

It runs and monitors:

• trader.py - waits for each bar close, loads data (from API or streamer CSVs), calls your signal function, runs risk checks, executes orders. Supports passive-first execution (--passive): posts a limit order at the best bid/ask, waits 10 seconds, crosses at market if unfilled. Handles position sync, flip logic, coin-specific size rounding, and graceful shutdown (flattens all on Ctrl+C).

• trade_monitor.py - polls positions every 60 seconds. Tracks account value (perps + spot), unrealized P&L, margin used, and drawdown from peak. Logs everything to positions.csv. If drawdown exceeds your threshold, it flattens all positions via market orders and writes a KILL_SWITCH file that stops the trader from re-opening positions. Delete the file to resume trading.

Same crash recovery and internet handling as the streamer supervisor. Checks every 30 seconds, restarts on crash, kills on hang (10 minutes no output).

Risk Engine

Pluggable risk checks that run before every trade. Built-in: max exposure, max position, daily loss limit, drawdown kill switch, concentration limit, max open positions. Write your own:

from risk import RiskEngine, max_exposure, max_drawdown

def no_weekend_trading(ctx):
    if ctx.now.weekday() >= 5:
        return RiskResult(allow=False, reason="Weekend")
    return RiskResult(allow=True)

engine = RiskEngine()
engine.add(max_exposure(500))
engine.add(max_drawdown(5.0))
engine.add(no_weekend_trading)

Or use CLI flags: --max-exposure 500 --max-position 100 --max-daily-loss 50

Execution

Two modes:

• Market orders (default): cross immediately with slippage tolerance

• Passive-first (--passive): post limit at best bid/ask, wait 10 seconds, cross if unfilled. Saves ~7 bps per round trip

The executor handles position sync, flip logic (close then reopen when switching sides), coin-specific size rounding (fetched from the exchange metadata), and graceful shutdown (flattens all on SIGINT/SIGTERM).

Share

Data

Three ways to get historical data:

• REST API: python download_candles.py --coins BTC,ETH --interval 4h --days 365 (paginates automatically, no auth)

• S3 tick archive: python download_s3.py --coins BTC,ETH --days 90 --interval 5m (millisecond fills, needs AWS CLI)

• WebSocket streamer: real-time bars via the streamer supervisor

Backtesting

Three strategy templates included. Details will be given in a separate post. Each downloads data, runs a backtest, and prints the commands to go live:

# Mean reversion: z-score fade
python connectors/hyperliquid/strategies/mean_reversion_template.py --days 60

# Buy the dip: vol-scaled entries on X% drops
python connectors/hyperliquid/strategies/buy_the_dip_template.py --threshold -5 --hold 48

# Feature alpha: scan 336 features, build composite
python feature_engine/example_hyperliquid_features.py --days 60
python connectors/hyperliquid/strategies/feature_alpha_template.py

Install

conda env create -f environment.yml && conda activate delphic-desk
# Or: pip install -r requirements.txt (needs TA-Lib C library)

echo 'export HYPERLIQUID_PRIVATE_KEY="0x..."' > infra/.env
echo 'export HYPERLIQUID_ACCOUNT_ADDRESS="0x..."' >> infra/.env

The repo includes a Getting Started guide (PDF) covering account setup, API wallets, funding, and testnet configuration.

Related posts: Building an Alpha Dataset from Hyperliquid (data pipeline + feature scan) and Strategy Templates for Hyperliquid (detailed backtesting guide).

Disclaimer: This is infrastructure, not investment advice. The strategy templates are for research. Whether a signal makes money depends on the signal, not the plumbing.

Refer a friend

The Strategy

Cross-sectional trend following is the core strategy behind most managed futures (CTA) funds. The idea goes back to the 1980s: rank a universe of instruments by the strength of their recent trend, go long the strongest, short the weakest. The strategy is market-neutral by construction - it doesn't care whether markets are going up or down, only which instruments are trending hardest relative to the others.

The version I test here uses exponentially weighted returns rather than simple lookback windows, blends three trend speeds to avoid overfitting to a single horizon, and normalizes everything by volatility so that a bond position is comparable to an energy position in risk terms. The full algorithm at each rebalance:

1. Trend signal: for each instrument, compute the exponentially weighted mean of daily returns at three speeds (halflife 10, 30, 120 days). This gives more weight to recent returns while capturing the full trend history

2. Vol-adjustment: divide each signal by the instrument's EWMA volatility (halflife 32 days). A 5% move in natural gas becomes comparable to a 0.3% move in bonds

3. Rank and blend: rank each speed cross-sectionally (percentile 0 to 1), then blend with weights 25% fast + 50% medium + 25% slow

4. Portfolio construction: long the top N instruments by blended score, short the bottom N

5. Position sizing: weight each position by inverse EWMA volatility within its leg (risk parity)

6. Execution: rebalance every H days, execute at the open

The Experiment

I apply this signal to four asset classes:

• Futures: 111 contracts (equities, rates, energy, metals, ags, FX), hourly data snapped at 09:00 ET, 2008-2026

• Crypto: 80 coins, hourly data resampled to daily, 2017-2026

• US Stocks: 2627 liquid equities (>$5M daily dollar volume), daily data, 2000-2026

• FX: 79 pairs, hourly data resampled to daily, 2010-2026

For each asset class, I rank instruments by liquidity and test at increasing universe sizes (top 10, top 20, ..., up to the full universe). At each size I sweep two parameters:

• N per leg: the number of instruments to long and short (futures/FX/crypto: 3, 5, 7, 10; stocks: 5, 10, 20, 50)

• Holding period H: how many days to hold before rebalancing (futures/FX/crypto: 5, 10, 15, 21d; stocks: 10, 21, 63d)

I select the (N, H) pair with the highest net Sharpe at each universe size. The cost models are asset-class specific: futures 5bps, stocks 3bps, FX 2bps, crypto 5bps per unit of turnover.

Total: 2897 instruments, 36 universe-size configurations, each optimized over multiple (N, H) pairs.

Related Research

• US Equity Cross-Sectional Mean Reversion - the opposite strategy on stocks: buying losers and selling winners. Where momentum fails on stocks, mean reversion thrives

• Crisis Alpha: Which Signals Thrive Under Each Macro Regime? - how different signals perform across bull, bear, and crisis environments

• The Delphic Fund - a live portfolio combining multiple tested strategies from this research into a single allocation

Subscribe now

Results

The headline finding: trend following performance varies dramatically by asset class. Futures and crypto show strong, persistent momentum that improves as you add more instruments. Stocks are marginal at best, requiring quarterly rebalancing to show any positive signal. FX is flat until you reach 60+ pairs. The relationship between universe size and performance is not linear - each asset class has a sweet spot beyond which adding more instruments dilutes the signal with illiquid noise.

The next chart breaks this down per asset class, showing gross and net Sharpe (top row) and average return per trade in basis points (bottom row) at each universe size tested.

Best Configuration per Asset Class

Why the Differences?

Futures trend because they are driven by macro forces that persist: central bank rate cycles, commodity supply shocks, currency policy divergences. When oil is trending up because of OPEC cuts, that cause doesn't reverse overnight. When bonds are selling off because a central bank is hiking, the hiking cycle lasts quarters. The trend continues because the underlying driver persists. Adding more futures to the cross-section adds more independent macro themes (energy, rates, metals, agriculture, FX), which is why diversification genuinely helps.

Crypto trends for a different reason: the asset class is dominated by momentum-chasing retail flow and narrative cycles (DeFi summer, NFT mania, AI tokens). When a coin starts running, the crowd piles in. The feedback loop is strong but unstable, which explains the high Sharpe paired with extreme drawdowns.

Stocks are different. Individual stock prices mean-revert at short horizons because market makers and statistical arbitrage desks provide liquidity by buying dips and selling rips. This microstructure-driven reversal works against the momentum signal. Stocks do trend at longer horizons (hence the marginal positive signal at 63-day holds), but the effect is weak and inconsistent.

FX is the weakest because currency trends are shorter-lived and more frequently interrupted by central bank intervention. The cross-section is also more homogeneous than futures - most FX pairs share common factors (USD strength, risk appetite), which limits the diversification benefit of adding more pairs.

Caveats

Every data point on the Sharpe vs Universe Size chart represents the best (N, H) configuration at that universe size - selected in-sample. This means the absolute Sharpe levels are optimistic. The relative ordering across asset classes is more reliable than the absolute numbers, because the same optimization procedure is applied everywhere. The data periods also differ: futures have 18 years, stocks 26 years, crypto only 9 years. Longer samples are more reliable.

Summary

• Futures are the strongest asset class for cross-sectional momentum, peaking at 80 instruments with net Sharpe 0.73. The signal is positive from 20 instruments onward

• Crypto shows strong momentum (net Sharpe 0.56 at 70 coins) but with large drawdowns (-41%), consistent with the high volatility of the asset class

• Stocks are weakly positive at best (net Sharpe 0.42) and only with quarterly rebalancing. The short-term mean reversion that dominates equity microstructure works against the momentum signal at shorter horizons

• FX momentum is marginal (net Sharpe 0.24), only emerging at 60+ pairs. FX trends exist but they are weaker and shorter-lived than in futures or crypto

• More instruments is not always better - each asset class has a peak beyond which adding illiquid instruments dilutes the signal

Share

Methodology

Signal: exponentially weighted mean return at halflife 10/30/120 days, divided by EWMA vol (halflife 32), rank-normalized cross-sectionally, blended 25/50/25. Position sizing: inverse-vol weighted within each leg. Execution: open-to-open. Futures snapped at 09:00 ET for cross-sectional alignment. FX and crypto resampled from hourly to daily. Stocks use split- and dividend-adjusted daily data. Returns clipped at +/-30% (futures/FX/crypto) and +/-50% (stocks). For each universe size, the (N, H) pair with the highest net Sharpe is selected.

Disclaimer: This is not investment advice. All results are backtested with optimal parameter selection at each universe size (in-sample). Past performance does not predict future results.

Subscribe now

All you need is a laptop, Python, and Jupyter Notebook. Three strategies on free daily data (SPY, TLT, GLD), 16 years of results. Run the notebook after the US close, enter your account size, get positions for the next day. None of the strategies are impressive on their own. That is the point. The ensemble is where the lesson is.

Subscribe now

Strategy 1: Momentum

If price is above where it was 126 days ago, go long. Otherwise stay flat. The simplest possible systematic strategy. Enter at the next open, hold to close.

Strategy 2: Mean Reversion (Fade Breakout on Bonds)

When price breaks above the 5-bar high, short it. When it breaks below the 5-bar low, buy it. This fights the trend, so it fails on equities but works on bonds where central bank anchoring and flight-to-quality flows reverse.

Strategy 3: Calendar Effect on Gold (Turn-of-Month)

Read more

On June 15 I started tracking 14 strategies live against the same backtest that produced Sharpe 1.64 over 14 years. Four weeks is not enough to judge a portfolio. But it is enough to see whether the machine works: do the signals fire on schedule, does the diversification hold up under real market moves, and is anything broken?

Short answer: the machine works. Longer answer below.

Portfolio: +0.38%. S&P 500: +2.09%. The portfolio underperformed in absolute terms. That is expected. It runs at 5.0% annualized volatility versus the S&P's 14.4%, roughly 3x less risk. The return per unit of risk is what matters, and at this early stage the portfolio is on track.

The more interesting number: the portfolio's worst drawdown was -0.97% versus the S&P's -3.17%. When equities sold off in late June, the portfolio barely moved. That is not luck. That is 14 strategies firing at different times in different markets, exactly as designed.

Related: The Delphic Fund | Weekly Tracker | Daily Dashboard | Technical Details

Subscribe now

Who Carried the Weight

8 strategies positive, 4 negative, 2 flat.

The biggest raw return was Leverage for the Long Run at +5.79%. But it only has a 1.2% portfolio weight (it is too volatile for more), so its actual contribution was just +0.07%.

The strategy that actually moved the needle was Pre-Holiday/FOMC. It returned +1.10% on a 21.6% weight, contributing +0.24% to the portfolio, more than any other strategy. This is inverse-volatility weighting doing its job: boring, low-vol strategies with modest returns contribute more than exciting, high-vol strategies with big returns.

The biggest drag was Turnaround Tuesday at -1.45%, dragging the portfolio by -0.17%.

What Fired and What Didn't

Never fired: BTC/Gold Dual Momentum, IBS+RSI Mean Reversion. This is normal. These strategies have strict entry conditions and they correctly stayed in cash. A strategy that never fires in a calm month is working as intended, its value shows up during the specific conditions it was designed for.

Always on: Risk-Weighted 3-Asset, All-Weather Portfolio, SPY/TLT/EEM Rotation, Leverage for the Long Run, Low-Vol Trend Follow, 7-ETF Rotation. These stay invested at all times and rebalance based on momentum or volatility signals. They are the backbone of the portfolio.

The Drawdown That Didn't Happen

The S&P dropped -3.17% from its peak during this period. The portfolio dropped -0.97%. That is a 3.3x reduction in drawdown.

This is the core promise of the portfolio: not higher returns, but dramatically lower drawdowns. Over 14 years of backtesting, the portfolio's worst drawdown is -11.0% versus the S&P's -55%. Four weeks of live data shows the same pattern holding.

Is the Backtest Lying?

The full backtest shows Sharpe 1.64, CAGR 6.3%, max drawdown -11.0%. The first 18 trading days produced +0.38%.

Too early to tell. 18 days is noise. But three things give me confidence the backtest is not lying:

1. Signals fire when expected. Every strategy has triggered at the right time with the right instrument. No implementation bugs found.

2. Drawdown protection works. The 3.3x drawdown reduction is consistent with the backtest's long-run ratio.

3. Correlation is real. The average pairwise correlation between strategies in the backtest is 0.16. In these 18 days, the diversification benefit showed up exactly when it mattered, during the late-June equity selloff.

The real test is 3-6 months. If the portfolio is still tracking the backtest by month three, the signal is credible. If it diverges, I will say so here.

What's Next

I am currently testing new strategy additions to the portfolio. Any strategy that passes the same robustness tests as the original 14 will be added and tracked live. If you have a strategy you think belongs here, drop it in the comments or send me a DM.

The weekly tracker has individual strategy tearsheets updated every Saturday. The daily dashboard shows current positions every morning.

Related

• The Delphic Fund - how the portfolio is built and why these 14 strategies

• How Much Leverage Before Ruin? - why low-vol portfolios can be levered safely

• Alpha Combinations - the mathematics of strategy diversification

Methodology

OOS period: 2026-06-15 to 2026-07-10 (18 trading days). Strategies tracked using the same code and parameters as the backtest. Inverse-volatility weights computed from the 252 trading days prior to OOS start and held fixed. No transaction costs applied. Past performance does not predict future results.

Cross-sectional mean reversion is a good example of quant research where the signal is easy to find but the real question is whether it survives execution. I tested up to 3,365 rolling liquidity-screened US stocks (3,204 at 5M, 3,263 at 30M/1H, 3,365 daily) across four timeframes, ~40 parameter combinations each, and 10+ years of intraday data to find out.

The Strategy

The canonical cross-sectional reversal. At each bar:

1. Compute the return over the last N bars for every stock in the universe

2. Rank all stocks from biggest loser to biggest winner

3. Long the K biggest losers, Short the K biggest winners

4. Equal weight within each leg, 1/K per position

5. Enter at the next bar's open, exit at the next bar's close. If a stock remains in the portfolio at the next rebalance, hold it (position carry) - no exit and re-entry, no cost on the held portion

No indicators, no machine learning, no signal filters. Intraday positions do not cross overnight. I tested this at four frequencies: 5-minute, 30-minute, hourly, and daily. At each, I swept K (portfolio size: 3 to 50) and N (signal lookback: 1 bar to 1 day), giving ~40 combinations per timeframe.

Everything below is a backtest. Stocks must pass a prior-year liquidity filter (>$5M median daily dollar volume, >$5 price): 3,204 at 5M, 3,263 at 30M/1H, 3,365 daily. Intraday ranges: 5M January 2016 to January 2026; 30M/1H January 2016 to February 2026. Daily: January 2000 to June 2026 (26 years).

The Results: All Four Timeframes

Read more

Objective: test whether the market behaviour of Musk-linked companies contains a repeatable dip-buying signal, or whether the apparent "Elon Effect" is just narrative fitted after the fact. Tesla is the long-history subject, SpaceX is the recent IPO case study, and a 4,734-listing IPO dataset provides the external baseline.

The Tesla Journey: 17 Crashes, +25,759%

Tesla went public on June 29, 2010 at $1.27 (split-adjusted). Sixteen years later, it trades above $400. That is a 25,759% total return.

But the path was anything but smooth. Tesla has suffered 17 drawdowns of 30% or more from its peak. Five of those exceeded 50%. The worst, from the November 2021 peak through January 2023, was a 74% crash that wiped out more than three-quarters of the stock's value.

Every single one of those crashes was followed by a recovery. That is the empirical form of the "Elon Effect" thesis: Musk-linked companies appear to combine unusually frequent large drawdowns with unusually strong post-drawdown recoveries. The question is not whether the story sounds plausible. The question is whether the signal remains visible after defining the events mechanically and comparing them with reasonable baselines.

Disclaimer: this is an empirical observation from historical data, not investment advice. Past drawdown recoveries do not guarantee future ones. Tesla and SpaceX are single stocks with concentrated risk. Position sizing, personal risk tolerance, and the possibility of permanent capital loss should all be considered before acting on any pattern described here.

Two Musk IPOs, One Pattern

Tesla's IPO day was a textbook hot debut: +25.7% open-to-close on day one. Our analysis of 4,734 IPOs shows the median stock closes 0.04% below its open. Tesla was in the top 5% of day-one movers.

What happened next also matched our data: Tesla lost 16.8% from its opening price by day 5, then spent the next two weeks recovering. Volume collapsed from 281 million shares on day one to 44% by day two, almost exactly our median of 38%.

SpaceX, sixteen years later, followed the same script: +7.3% on day one, volume collapsing 51% by day two, a peak at $226 on day four, then a 35% crash to $147 by day eleven.

The pattern is similar because the mechanism is similar: speculative demand overshoots on the first day, liquidity evaporates, and the stock reprices. The only question is what happens after the initial post-IPO crash, and more broadly, after every subsequent crash. In the IPO baseline, the average stock drifts lower for weeks. Tesla also crashed, but its subsequent path differs from the baseline because repeated recovery, not the initial fall, is the dominant feature.

The Drawdown Recovery Machine

This is the key chart. The drawdown threshold is treated as the event marker and the following six-month return is the response variable. When Tesla enters a 30% drawdown, the mean forward return over the next six months is +27.5%, with a 77% win rate across 64 episodes. At 50% drawdowns, the response is larger: +84.3% mean return, 88% win rate.

How does this compare to other mega-cap tech stocks?

Tesla's recovery is competitive with the best (Meta's +41% from 30% drawdowns is the strongest in the group), but what sets Tesla apart is the frequency. Tesla has had 17 drawdowns of 30% or more. Apple has had 6. Google has had 2. Tesla gives you more opportunities to buy the dip, and each opportunity is roughly as profitable as the competition.

The deeper the dip, the bigger the recovery. Read as a signal-processing problem, drawdown depth is the input feature and forward return is the target. The relationship is noisy, but the fitted direction is clear: deeper drawdowns carry higher subsequent returns in this sample.

The Dip-Buying Alpha

Tesla has had 218 trading days with a loss exceeding 5%. That is 12% of all trading days, compared to just 16 for the S&P 500 over the same period. This alone tells you that Tesla is a fundamentally different asset: annualized volatility of 57% versus 17% for SPY.

There are two ways to parameterise the dip. First, single-day drops: what happens after a >5% or >10% daily loss? Second, cumulative drawdowns from the all-time high: what happens when the stock has fallen 15% or 30% from its peak?

Single-day drops show a modest edge over baseline, growing with severity. But what about sustained drawdowns from the all-time high? This is where the picture gets more nuanced.

The 15% drawdown shows negative excess at the 20-day horizon. This is the falling-knife problem: when Tesla has dropped 15% from its peak, it often keeps falling. In 40% of cases, a 15% drawdown continued to 25% or worse within 20 days. You are buying mid-crash, not at the bottom.

An apparent contradiction: daily drops of >5% show positive excess returns, yet entering at a 15% drawdown shows negative excess. The explanation is that these are not the same feature. 82% of >5% daily drops happen when the stock is already in a drawdown of 15% or more. The median drawdown at the time of a >5% daily loss is -29%. The daily-drop signal is therefore not detecting the start of a crash; it is detecting a volatility spike inside an already developed drawdown. A 15% drawdown entry, by contrast, is the first day the stock crosses that level. It is early in the crash, before the floor has been established.

At 30% drawdown, the excess turns positive across all horizons. By that point, most of the selling pressure has exhausted itself and the recovery begins. The real dip-buying alpha is not in reacting to daily drops; it is in waiting for the drawdown to reach a level where the odds genuinely shift in your favour. Tesla's annualized volatility is 57%, more than 3x the S&P 500. That volatility is what creates these opportunities in the first place.

The Strategy section below treats this as a small classification and timing problem. It tests 19 strategy variants across drawdown levels, holding periods, retest filters, and combined signals, with annualized Sharpe, Calmar, win rates, and trade counts. One combined variant has a 79% win rate across 14 trades.

Read more

Multi-Asset Dip (post): 23 trades on ES, DAX, Nikkei, Crude Oil, Gold. 21-day hold after a threshold breach. 12m total: +10.2%.

S&P 500 Sector Dip (post): 14 trades on 11 sector ETFs. Vol-normalised 2.5-sigma threshold, 10-day hold. 12m total: +4.1%.

S&P 500 Stock Pullback (post): Individual stock dip buying across 503+ S&P 500 stocks. Vol-normalised 2.5-sigma threshold, severity/pre-vol filter, 10-day hold. Live tracked: 4 closed, 34 open, total PnL: +14.3%.

Opening Dip (post): 77 intraday trades on 14 futures. Buy the gap-up morning flush after a down day, hold to RTH close. 12m total: -51 bps.

Crypto Dip (post): Drawdown-based dip buying on 44 crypto perpetuals across 4 timeframes. Vol-scaled, net of 10 bps. 12m total (3d timeframe): +4.0%.

Paid subscribers get the individual equity curves for each strategy below, with per-instrument breakdown and recent trades.

Read more

The answer is yes, but only for the deeper dips. Shallow dips get eaten by trading costs. The numbers below use a rolling 61-coin Binance Futures universe through 2026-06-02.

Related: Buying the dip on S&P 500 stocks | Multi-asset dip strategy | The Delphic Fund | The Alpha Atlas

Every strategy in this study is defined by three parameters:

• Heartbeat (5 minutes): How often the strategy checks for signals. I use 5-minute candle data from Binance Futures, so the strategy evaluates every 5 minutes whether a coin has dipped enough.

• Lookback (1h / 4h / 1 day / 3 days): The window for detecting the rolling high. A "1-day lookback" means the strategy tracks each coin's highest price over the last 288 five-minute bars (= 24 hours). When the current price drops far enough below that high, it is a dip signal.

• Hold horizon (same as lookback): How long you hold after buying the dip. If the lookback is 1 day, you hold for 1 day. If 4 hours, you hold for 4 hours. This keeps the trade duration matched to the timeframe of the dip.

On top of these, position sizes are scaled inversely to each coin's trailing 1-day realized volatility. A 5% dip in a low-vol coin gets a bigger position than the same dip in a high-vol coin. This prevents the strategy from loading up during volatile regimes when dips tend to be trend continuations rather than reversals.

Every result below is a diversified portfolio across all eligible coins with data through 2026-06-02. New coins enter the portfolio when they have enough history for the lookback and volatility estimate, so the universe grows over time - just as it would for a real trader. Daily P&L is divided by the number of eligible coins that day, including coins with no signal.

Portfolio Performance (Raw, Before Risk Management)

Each panel shows the best dip threshold for that lookback/hold combination, with no position cap and no stop-loss. This is the raw signal - risk management comes later in the paid section. Green is gross, red is net of 10bps costs:

Which Strategies Survive Costs?

Annualized Sharpe ratios for every lookback/threshold combination, net of costs. Deeper dips on longer lookbacks produce the most robust edges. Shallow dips get destroyed by execution costs:

Key Results

• Best raw portfolio: Lookback 1 day, Hold 1 day, dip at -12%. Sharpe 1.90 net of costs. 61 coins generated at least one qualifying trade inside a 61-coin eligible universe.

• Destroyed by costs: Lookback 1h/-2%, Lookback 4h/-3%, Lookback 1 day/-5% go negative after 10bps round-trip.

• Robust in the basic tests: The best strategy has t-stat 2.9 (one-sided p = 0.0019). Bootstrap 95% CI for Sharpe: [0.7, 3.3]. Treat this as a sanity check, not a final proof, because crypto coins are correlated and the thresholds were selected after testing.

• Consistent across time: Sharpe of 1.8 in the first half, 2.6 in the second half. 76% of months profitable.

How to Trade This (Practical Guide + Risk Management)

This is educational analysis, not investment advice. Backtested results do not guarantee future performance. Crypto futures are highly volatile and you can lose more than your initial investment. Do your own research and only trade with capital you can afford to lose.

Read more

The Delphic Fund: Technical Details

5-year equity curves and full backtest metrics for all 14 strategies.

Last updated: August 01, 2026

Each strategy below shows a 5-year equity curve and a side-by-side comparison of recent vs full backtest metrics. Strategies are sorted by full-backtest Sharpe ratio.

Strategy Tearsheets

MonthEnd Boundary

Metric: Sharpe | Last 5 Years: 1.06 | Full Backtest: 1.37

Metric: Return | Last 5 Years: +13.3% | Full Backtest: -

Metric: CAGR | Last 5 Years: 2.5% | Full Backtest: 2.8%

Metric: Volatility | Last 5 Years: 2.4% | Full Backtest: 2.0%

Metric: Max Drawdown | Last 5 Years: -3.2% | Full Backtest: -3.2%

Metric: Win Rate | Last 5 Years: 55% | Full Backtest: 57%

Metric: % Time Invested | Last 5 Years: 37% | Full Backtest: 37%

Metric: Min Sharpe (param test) | Last 5 Years: 0.75

Metric: Stability Score | Last 5 Years: 15.5

Metric: Full Period | Last 5 Years: 2003-01-02 to 2026-07-31

Last 6 months: Feb 26: +0.1% | Mar 26: +0.1% | Apr 26: -0.2% | May 26: +0.3% | Jun 26: +0.5% | Jul 26: -0.2%

BTC Gold DualMom

Metric: Sharpe | Last 5 Years: 0.84 | Full Backtest: 1.18

Metric: Return | Last 5 Years: +241.5% | Full Backtest: -

Metric: CAGR | Last 5 Years: 27.9% | Full Backtest: 70.6%

Metric: Volatility | Last 5 Years: 37.6% | Full Backtest: 60.0%

Metric: Max Drawdown | Last 5 Years: -42.7% | Full Backtest: -75.0%

Metric: Win Rate | Last 5 Years: 53% | Full Backtest: 53%

Metric: % Time Invested | Last 5 Years: 81% | Full Backtest: 82%

Metric: Min Sharpe (param test) | Last 5 Years: 1.04

Metric: Stability Score | Last 5 Years: 75.4

Metric: Full Period | Last 5 Years: 2013-06-28 to 2026-07-24

Last 6 months: Feb 26: +8.7% | Mar 26: -17.2% | Apr 26: -0.2% | May 26: -8.1% | Jun 26: -9.1% | Jul 26: +0.0%

RiskWeighted 3Asset

Metric: Sharpe | Last 5 Years: 0.72 | Full Backtest: 1.04

Metric: Return | Last 5 Years: +32.9% | Full Backtest: -

Metric: CAGR | Last 5 Years: 5.9% | Full Backtest: 7.9%

Metric: Volatility | Last 5 Years: 8.4% | Full Backtest: 7.7%

Metric: Max Drawdown | Last 5 Years: -16.4% | Full Backtest: -16.4%

Metric: Win Rate | Last 5 Years: 54% | Full Backtest: 55%

Metric: % Time Invested | Last 5 Years: 100% | Full Backtest: 100%

Metric: Min Sharpe (param test) | Last 5 Years: 1.01

Metric: Stability Score | Last 5 Years: 44.1

Metric: Full Period | Last 5 Years: 2005-02-18 to 2026-07-31

Last 6 months: Feb 26: +3.1% | Mar 26: -5.3% | Apr 26: +2.3% | May 26: +1.2% | Jun 26: -1.3% | Jul 26: -1.6%

PreHoliday FOMC

Metric: Sharpe | Last 5 Years: 0.73 | Full Backtest: 0.89

Metric: Return | Last 5 Years: +9.3% | Full Backtest: -

Metric: CAGR | Last 5 Years: 1.8% | Full Backtest: 1.7%

Metric: Volatility | Last 5 Years: 2.5% | Full Backtest: 1.9%

Metric: Max Drawdown | Last 5 Years: -2.1% | Full Backtest: -5.9%

Metric: Win Rate | Last 5 Years: 61% | Full Backtest: 60%

Metric: % Time Invested | Last 5 Years: 23% | Full Backtest: 22%

Metric: Min Sharpe (param test) | Last 5 Years: 0.48

Metric: Stability Score | Last 5 Years: 4.7

Metric: Full Period | Last 5 Years: 2004-11-18 to 2026-07-31

Last 6 months: Feb 26: -0.4% | Mar 26: +0.2% | Apr 26: -0.3% | May 26: +0.0% | Jun 26: +0.7% | Jul 26: +0.9%

WealthMgr Rebal

Metric: Sharpe | Last 5 Years: 1.16 | Full Backtest: 0.88

Metric: Return | Last 5 Years: +60.1% | Full Backtest: -

Metric: CAGR | Last 5 Years: 9.9% | Full Backtest: 8.0%

Metric: Volatility | Last 5 Years: 8.4% | Full Backtest: 9.2%

Metric: Max Drawdown | Last 5 Years: -10.3% | Full Backtest: -18.0%

Metric: Win Rate | Last 5 Years: 56% | Full Backtest: 57%

Metric: % Time Invested | Last 5 Years: 28% | Full Backtest: 28%

Metric: Min Sharpe (param test) | Last 5 Years: 0.50

Metric: Stability Score | Last 5 Years: 5.1

Metric: Full Period | Last 5 Years: 2003-01-02 to 2026-07-31

Last 6 months: Feb 26: +0.5% | Mar 26: +0.3% | Apr 26: -1.3% | May 26: +1.8% | Jun 26: +0.3% | Jul 26: -1.4%

SPY TLT EEM Rotation

Metric: Sharpe | Last 5 Years: 0.43 | Full Backtest: 0.76

Metric: Return | Last 5 Years: +36.1% | Full Backtest: -

Metric: CAGR | Last 5 Years: 6.4% | Full Backtest: 13.6%

Metric: Volatility | Last 5 Years: 18.4% | Full Backtest: 19.3%

Metric: Max Drawdown | Last 5 Years: -39.2% | Full Backtest: -39.2%

Metric: Win Rate | Last 5 Years: 52% | Full Backtest: 54%

Metric: % Time Invested | Last 5 Years: 100% | Full Backtest: 100%

Metric: Min Sharpe (param test) | Last 5 Years: 0.58

Metric: Stability Score | Last 5 Years: 8.6

Metric: Full Period | Last 5 Years: 2004-01-02 to 2026-07-31

Last 6 months: Feb 26: +5.9% | Mar 26: -9.3% | Apr 26: +12.7% | May 26: +7.2% | Jun 26: -1.0% | Jul 26: -6.3%

AllWeather 333

Metric: Sharpe | Last 5 Years: 0.65 | Full Backtest: 0.74

Metric: Return | Last 5 Years: +39.5% | Full Backtest: -

Metric: CAGR | Last 5 Years: 6.9% | Full Backtest: 6.8%

Metric: Volatility | Last 5 Years: 11.4% | Full Backtest: 10.0%

Metric: Max Drawdown | Last 5 Years: -23.7% | Full Backtest: -23.7%

Metric: Win Rate | Last 5 Years: 55% | Full Backtest: 55%

Metric: % Time Invested | Last 5 Years: 100% | Full Backtest: 100%

Metric: Min Sharpe (param test) | Last 5 Years: 0.65

Metric: Stability Score | Last 5 Years: 72.2

Metric: Full Period | Last 5 Years: 2012-01-24 to 2026-07-31

Last 6 months: Feb 26: +4.8% | Mar 26: -6.9% | Apr 26: +2.2% | May 26: +1.1% | Jun 26: -3.8% | Jul 26: -1.1%

Turnaround Tuesday

Metric: Sharpe | Last 5 Years: 0.58 | Full Backtest: 0.72

Metric: Return | Last 5 Years: +13.3% | Full Backtest: -

Metric: CAGR | Last 5 Years: 2.5% | Full Backtest: 4.6%

Metric: Volatility | Last 5 Years: 4.5% | Full Backtest: 6.5%

Metric: Max Drawdown | Last 5 Years: -4.2% | Full Backtest: -11.6%

Metric: Win Rate | Last 5 Years: 49% | Full Backtest: 57%

Metric: % Time Invested | Last 5 Years: 7% | Full Backtest: 8%

Metric: Min Sharpe (param test) | Last 5 Years: 0.70

Metric: Stability Score | Last 5 Years: 48.7

Metric: Full Period | Last 5 Years: 2000-01-03 to 2026-07-31

Last 6 months: Feb 26: +0.7% | Mar 26: +2.9% | Apr 26: -0.7% | May 26: +0.1% | Jun 26: -1.5% | Jul 26: +1.2%

LowVol TrendFollow

Metric: Sharpe | Last 5 Years: 0.88 | Full Backtest: 0.70

Metric: Return | Last 5 Years: +59.9% | Full Backtest: -

Metric: CAGR | Last 5 Years: 9.9% | Full Backtest: 7.3%

Metric: Volatility | Last 5 Years: 11.4% | Full Backtest: 10.9%

Metric: Max Drawdown | Last 5 Years: -18.8% | Full Backtest: -20.7%

Metric: Win Rate | Last 5 Years: 56% | Full Backtest: 56%

Metric: % Time Invested | Last 5 Years: 78% | Full Backtest: 71%

Metric: Min Sharpe (param test) | Last 5 Years: 0.63

Metric: Stability Score | Last 5 Years: 80.8

Metric: Full Period | Last 5 Years: 2000-01-03 to 2026-07-31

Last 6 months: Feb 26: -0.9% | Mar 26: -5.2% | Apr 26: +6.3% | May 26: +5.3% | Jun 26: -1.0% | Jul 26: +0.0%

Leverage LongRun

Metric: Sharpe | Last 5 Years: 0.66 | Full Backtest: 0.70

Metric: Return | Last 5 Years: +127.3% | Full Backtest: -

Metric: CAGR | Last 5 Years: 17.9% | Full Backtest: 20.2%

Metric: Volatility | Last 5 Years: 34.0% | Full Backtest: 35.2%

Metric: Max Drawdown | Last 5 Years: -48.9% | Full Backtest: -51.2%

Metric: Win Rate | Last 5 Years: 55% | Full Backtest: 55%

Metric: % Time Invested | Last 5 Years: 78% | Full Backtest: 82%

Metric: Min Sharpe (param test) | Last 5 Years: 0.72

Metric: Stability Score | Last 5 Years: 87.8

Metric: Full Period | Last 5 Years: 2010-01-04 to 2026-07-31

Last 6 months: Feb 26: -3.6% | Mar 26: -15.8% | Apr 26: +19.1% | May 26: +15.5% | Jun 26: -4.7% | Jul 26: -1.1%

Holiday Energy

Metric: Sharpe | Last 5 Years: 0.75 | Full Backtest: 0.68

Metric: Return | Last 5 Years: +84.0% | Full Backtest: -

Metric: CAGR | Last 5 Years: 13.0% | Full Backtest: 11.8%

Metric: Volatility | Last 5 Years: 18.6% | Full Backtest: 19.1%

Metric: Max Drawdown | Last 5 Years: -23.9% | Full Backtest: -52.5%

Metric: Win Rate | Last 5 Years: 54% | Full Backtest: 54%

Metric: % Time Invested | Last 5 Years: 34% | Full Backtest: 31%

Metric: Min Sharpe (param test) | Last 5 Years: 0.53

Metric: Stability Score | Last 5 Years: 7.3

Metric: Full Period | Last 5 Years: 2008-02-27 to 2026-07-31

Last 6 months: Feb 26: -0.6% | Mar 26: +0.0% | Apr 26: +0.0% | May 26: -6.1% | Jun 26: -2.9% | Jul 26: -3.3%

7ETF Rotation

Metric: Sharpe | Last 5 Years: 0.34 | Full Backtest: 0.66

Metric: Return | Last 5 Years: +22.9% | Full Backtest: -

Metric: CAGR | Last 5 Years: 4.2% | Full Backtest: 9.6%

Metric: Volatility | Last 5 Years: 15.5% | Full Backtest: 15.7%

Metric: Max Drawdown | Last 5 Years: -20.4% | Full Backtest: -26.8%

Metric: Win Rate | Last 5 Years: 58% | Full Backtest: 57%

Metric: % Time Invested | Last 5 Years: 96% | Full Backtest: 95%

Metric: Min Sharpe (param test) | Last 5 Years: 0.43

Metric: Stability Score | Last 5 Years: 4.0

Metric: Full Period | Last 5 Years: 2008-01-02 to 2026-07-31

Last 6 months: Feb 26: +4.6% | Mar 26: -9.3% | Apr 26: +0.3% | May 26: +10.6% | Jun 26: -0.1% | Jul 26: -6.6%

IBS RSI MeanRev

Metric: Sharpe | Last 5 Years: 0.29 | Full Backtest: 0.44

Metric: Return | Last 5 Years: +12.9% | Full Backtest: -

Metric: CAGR | Last 5 Years: 2.5% | Full Backtest: 4.6%

Metric: Volatility | Last 5 Years: 10.2% | Full Backtest: 11.8%

Metric: Max Drawdown | Last 5 Years: -17.1% | Full Backtest: -28.8%

Metric: Win Rate | Last 5 Years: 55% | Full Backtest: 55%

Metric: % Time Invested | Last 5 Years: 19% | Full Backtest: 19%

Metric: Min Sharpe (param test) | Last 5 Years: ?

Metric: Stability Score | Last 5 Years: ?

Metric: Full Period | Last 5 Years: 2000-01-03 to 2026-07-31

Last 6 months: Feb 26: +0.5% | Mar 26: +0.4% | Apr 26: +0.0% | May 26: +0.0% | Jun 26: +1.7% | Jul 26: +0.8%

Futures Dip

Metric: Sharpe | Last 5 Years: 0.53 | Full Backtest: 0.36

Metric: Return | Last 5 Years: +20.9% | Full Backtest: -

Metric: CAGR | Last 5 Years: 3.9% | Full Backtest: 3.2%

Metric: Volatility | Last 5 Years: 6.1% | Full Backtest: 7.9%

Metric: Max Drawdown | Last 5 Years: -8.5% | Full Backtest: -22.2%

Metric: Win Rate | Last 5 Years: 54% | Full Backtest: 54%

Metric: % Time Invested | Last 5 Years: 34% | Full Backtest: 33%

Metric: Min Sharpe (param test) | Last 5 Years: 0.10

Metric: Stability Score | Last 5 Years: 1.5

Metric: Full Period | Last 5 Years: 2008-01-01 to 2026-07-31

Last 6 months: Feb 26: +0.8% | Mar 26: +7.5% | Apr 26: +0.7% | May 26: +3.1% | Jun 26: -2.9% | Jul 26: -0.5%

Everything below is a backtest. These are historical results, not live returns. Live tracking started in June 2026 and is updated every Saturday in the Weekly Tracker.

Important: These are my independent implementations of published strategies, built from their descriptions. My versions may differ from the originals in ways that matter. Different data, different parameter choices, different execution assumptions. The original authors have not endorsed or verified my implementations. When in doubt, read the original. The purpose of the first few weeks of live tracking is to identify implementation errors and assess whether execution at the assumed prices is viable in practice.

See the technical details for diversification mathematics, exact rules and parameters for each strategy, individual equity curves, and weekly signal updates.

Related: Weekly Tracker (live results) | Technical Details | New to trading? Start here | Savings vs investing backtest | How much leverage before ruin? | The Alpha Atlas

Refer a friend

The Portfolio

: Risk-adjusted return (Sharpe) | Backtest (2005-2026): 1.31

: Annual return | Backtest (2005-2026): 3.6%

: Worst drawdown | Backtest (2005-2026): -11.2%

14 strategies, inverse-volatility weighted over a common 2005-2026 window. Strategies with smoother daily returns receive more capital; high-volatility strategies such as BTC/Gold and leveraged equity receive small weights.

These are backtest results, not live returns. The backtest is the hypothesis. The weekly updates are the experiment.

Why These 14?

I started with a few dozen strategies. Most were cut — dead edges, fragile parameters, look-ahead bias, or unreliable data. The survivors all remained profitable when I varied their key parameters across a reasonable range. This list is not final. I am actively screening new strategies and will add any that pass the same robustness tests.

The average pairwise correlation between the 14 strategies is 0.16. Most are nearly uncorrelated, which is why the portfolio Sharpe (1.31) is higher than any individual strategy. Compared to SPY buy-and-hold over the same period: the portfolio's worst drawdown (-11.2%) is one-fifth of SPY's (-55%), at the cost of lower absolute returns (3.6% vs 10.9%).

The low absolute return is a feature, not a bug. The portfolio runs at ~3% annualized volatility, roughly one-seventh of the S&P 500's ~19%. At 2x leverage the portfolio matches SPY's 11% CAGR with 6% volatility and a -19% max drawdown, compared to SPY's 19% volatility and -55% drawdown. At 3x leverage it beats SPY outright: 15% CAGR, 10% volatility, -27% max drawdown. The chart below shows all three alongside SPY.

Parameter sensitivity matters, but it is not magic. A strategy that only works at one exact setting is suspect; a strategy that remains positive across nearby settings is more credible. The table below shows the backtest Sharpe alongside the worst-case Sharpe from parameter sweeps.

The 14 Strategies

Sorted by Sharpe ratio.

Strategy: Month-End Boundary | Source: Beyond Passive | Sharpe: 1.37 | Weight: 19.2%

Strategy: BTC/Gold Dual Momentum | Source: Valuelytica | Sharpe: 1.21 | Weight: 1.1%

Strategy: Risk-Weighted 3-Asset | Source: Beyond Passive | Sharpe: 1.05 | Weight: 3.9%

Strategy: Futures Dip | Source: Delphic Alpha | Sharpe: 1.03 | Weight: 18.9%

Strategy: Wealth Manager Rebalancing | Source: Robot Wealth | Sharpe: 0.89 | Weight: 5.5%

Strategy: SPY/TLT/EEM Rotation | Source: Quantified Strategies | Sharpe: 0.78 | Weight: 1.9%

Strategy: All-Weather Portfolio | Source: Dalio / Bridgewater | Sharpe: 0.77 | Weight: 2.8%

Strategy: Turnaround Tuesday | Source: Calendar anomaly (Cross 1973, French 1980) | Sharpe: 0.72 | Weight: 16.4%

Strategy: Low-Vol Trend Following | Source: Faber (2007) | Sharpe: 0.70 | Weight: 3.0%

Strategy: Leverage for the Long Run | Source: Hedgefundie (Bogleheads) | Sharpe: 0.70 | Weight: 1.0%

Strategy: Holiday Energy | Source: TradeQuantiX | Sharpe: 0.69 | Weight: 2.3%

Strategy: 7-ETF Rotation | Source: Antonacci (2014) | Sharpe: 0.68 | Weight: 1.9%

Strategy: Pre-Holiday/FOMC | Source: Beyond Passive | Sharpe: 0.61 | Weight: 16.2%

Strategy: IBS + RSI Mean Reversion | Source: Connors & Alvarez (2009) | Sharpe: 0.44 | Weight: 5.8%

How The Portfolio Works

Each strategy gets a fixed capital allocation based on inverse volatility: quieter strategies get more capital, volatile ones get less. The weights in the table above are recomputed quarterly using trailing 1-year daily volatility.

There is no rebalancing when a strategy fires. Each strategy operates independently. When a strategy signals "buy", it uses its allocated capital. When it signals "cash", that capital earns nothing. The portfolio return on any given day is the weighted sum of all 14 strategy returns.

In practice, on most days only a few strategies are active. Calendar strategies (Turnaround Tuesday, Month-End, Pre-Holiday) fire on specific dates. Trend strategies (Low-Vol, Leverage) are either fully invested or fully in cash. Rotation strategies (7-ETF, SPY/TLT/EEM) are always invested in something. Dip-buying strategies (Futures Dip, IBS+RSI) only fire after large declines.

All strategies currently use close-to-close returns. This assumes you can execute at or near the closing price, which is realistic for liquid ETFs but not perfect. Over time I plan to move to more realistic execution assumptions (e.g. open-to-close, VWAP fills, slippage models) and will report the impact on performance as I do.

Signal timing: Known days in advance | Strategies: Month-End Boundary, Pre-Holiday/FOMC, Turnaround Tuesday, Holiday Energy

Signal timing: Known intraday | Strategies: Low-Vol Trend, Leverage for the Long Run, All-Weather, Risk-Weighted 3-Asset, BTC/Gold Dual Momentum, 7-ETF Rotation, SPY/TLT/EEM Rotation, Wealth Mgr Rebalancing

Signal timing: Needs today's close | Strategies: IBS + RSI Mean Reversion, Futures Dip

How To Do This Yourself

Most of these strategies do not require you to sit in front of a screen. You need a brokerage account and some of the following ETFs: VTI, SPY, TLT, GLD, URTH, QQQ, IWM, EFA, EEM. Futures Dip requires 1-minute futures data.

Level: Just a calendar | What you need: Phone alarm, buy/sell on specific dates | Strategies: Month-End Boundary, Pre-Holiday/FOMC, Turnaround Tuesday, Holiday Energy

Level: A spreadsheet | What you need: One formula per month in Excel | Strategies: Risk-Weighted 3-Asset, All-Weather, SPY/TLT/EEM Rotation, BTC/Gold Dual Momentum, 7-ETF Rotation, Wealth Manager Rebalancing, Low-Vol Trend Following, Leverage for the Long Run, IBS + RSI Mean Reversion

Level: A Python script | What you need: Run a script once per day | Strategies: Futures Dip

Share

What Happens Next

Starting this week, I track all 14 strategies and the combined portfolio in real time. Every week I publish:

• How the portfolio is actually doing right now

• Which strategies are active and what they are signaling

• Whether any strategy has started breaking down

If a strategy stops working, it will show up in the data and I will flag it. The point is transparency.

The detailed tearsheets have individual equity curves, exact rules and parameters, and the raw data files.

More Research From Delphic Alpha

If you found this interesting, here is some of the research behind the portfolio:

• The Intraday Alpha Atlas — 12 alphas tested across 5 asset classes on 5-minute bars. One signal was positive in every single market.

• The Daily Stock Alpha Atlas — 30 alphas mapped across 1,500 US stocks. Exotic signals dominate at the daily horizon.

• The Atlas Monitor — 1,000+ alpha signals tracked weekly. Which are working, which are decaying, which just died.

• Alpha Combinations — 55% of alpha pairs destroy value when combined. The math of why diversification works in this portfolio.

• Buying the Dip — 10,595 dip events across 11 S&P 500 sectors. The evidence behind the Futures Dip strategy.

• The Anatomy of an IPO — 4,734 first trading days. What actually happens after a stock starts trading.

The Fine Print

This is not investment advice. These are backtested results from strategies published by other people on the internet. My implementations may differ from the originals. Past performance does not predict future results. The parameter sensitivity test helps but does not guarantee future performance.

Credits

None of these strategies are mine except Futures Dip. All implementations are mine. All mistakes are mine.

Substack authors:

• Beyond Passive Investing - Month-End Boundary, Pre-Holiday/FOMC, Risk-Weighted 3-Asset

• Quantified Strategies - SPY/TLT/EEM Rotation

• Valuelytica - BTC/Gold Dual Momentum

• TradeQuantiX - Holiday Energy

• Robot Wealth - Wealth Manager Rebalancing

• Delphic Alpha - Futures Dip

Academic / practitioner sources:

• Cross (1973), French (1980) - Turnaround Tuesday (day-of-week effect)

• Ray Dalio / Bridgewater - All-Weather Portfolio

• Meb Faber, A Quantitative Approach to Tactical Asset Allocation (2007) - Low-Vol Trend Following

• Gary Antonacci, Dual Momentum Investing (2014) - 7-ETF Rotation

• Hedgefundie (Bogleheads, 2019) - Leverage for the Long Run

• Larry Connors & Cesar Alvarez, Short Term Trading Strategies That Work (2009) - IBS + RSI Mean Reversion

Subscribe now

When price dips 100 basis points below the regular-hours open within the first 45 minutes, buy and hold to the close. Sharpe 1.…

Read more

This post does it differently. I start with all 820 pairwise combinations of 41 crypto perpetual futures. No pairs are pre-selected. A composite scoring method picks 15 pairs from in-sample data, then those pairs are tested on the next period's out-of-sample data. Everything is lagged: the pairs I trade were chosen before the test window began.

Bottom line: +70% portfolio return at 8 bps (34% annualized, Sharpe 1.7), +89% at 4 bps (42% annualized, Sharpe 1.9). 30 coins traded, 67% of selected pairs profitable. No look-ahead.

Walk-Forward Design

I use a lagged walk-forward. 6 months in-sample, 3 months out-of-sample, 3 month step. The pairs picked in fold N get tested in fold N+1's OOS window, not the same fold. Fold 0 only provides the first selection, giving 5 OOS windows.

Starting universe: 820 pairs from 41 coins (all Binance/Hyperliquid overlap, excluding dead tickers). No pre-filtering. The composite scoring does all the selection from IS data alone. Quote coin cap of 2 pairs max. Top 15 selected per fold. Entry z=2.0, exit z=0.2, 7-day OLS hedge, 8 bps round-trip costs.

All 5 OOS folds are included in portfolio metrics. No folds excluded.

Picking Pairs: What Works and What Doesn't

The obvious approach is to pick pairs with the highest in-sample Sharpe. I tested 18 IS metrics as predictors of OOS Sharpe. None of them work. Best R-squared: 0.037.

But two things do carry forward: variance ratio (how mean-reverting the spread is) and skew (whether losses are fat-tailed). A composite of low VR, low skew, and improving recent Sharpe splits pairs into clear tiers:

Bottom quintile: 58% hit rate, avg per-pair Sharpe 0.18. Top quintile: 82% hit rate, avg per-pair Sharpe 1.07. The score can't predict exactly how good a pair will be, but it clearly separates likely winners from likely losers. Note: these are individual pair Sharpes. The portfolio Sharpe is higher (~1.7) because running ~7 pairs simultaneously provides diversification.

I also tested other scoring methods: IS Sharpe alone (69% hit rate, worst fold negative), stability scoring with 3 overlapping sub-windows (best per-pair Sharpe but weakest fold barely breaks even), and VR alone (72% hit rate but most mean-reverting pairs aren't always the most profitable). The composite of 50% skew, 30% VR, 20% recency gave the most balanced results and is what I use below.

Results

At 8 bps round-trip (taker fees): +70% portfolio return (34% annualized), Sharpe 1.7, max drawdown -14%. At 4 bps (maker fills): +89% portfolio return (42% annualized), Sharpe 1.9, max drawdown -20%. ~7 concurrent positions, 30+ coins, all 5 folds included.

Share

Per-Fold Breakdown

3 of 5 folds profitable. Fold 1 (-18% raw) and fold 3 (-4% raw) lose money. Fold 5 (Dec 2025 to Mar 2026) is the strongest at +213% raw.

Which Pairs Survive?

DYDX/STRK (+43%), AVAX/WLD (+37%), NEAR/ORDI (+36%), DOT/DYDX (+33%) lead. 30 unique coins traded across all folds. No single pair dominates, pairs rotate fold to fold.

What Didn't Work

Kalman filter (overtrades at 5-min), BTC regime filters (no link between BTC momentum and pair PnL), persistence filters (requiring pairs in consecutive IS top lists, too strict), nested IS validation (split IS in half, didn't improve hit rate), IS Sharpe as a ranking signal (zero correlation with OOS Sharpe).

Honest Assessment

• No pre-filtering: all 820 pairs compete in every fold. The composite scoring does all selection from IS data. No look-ahead.

• Fold 1 is weak: cold-start with limited IS history. Excluded from headline metrics.

• Fold 3 is near-flat: Jun to Sep 2025 is thin across every method I tested. The edge is regime-dependent.

• IS can't rank pairs: I can tell which pairs will make money but not which will make the most. Equal-weight is the right call.

• 30 coins: the scoring naturally diversifies across coins without needing to hand-pick a universe.

Subscribe now

Read more

Price rallies 100 bps from the open, pulls back, retests the high, and you buy for continuation. Sounds classic. The raw…

Read more

TL;DR: "Ensembling" in systematic trading is not one technique — it's at least seven distinct operations that happen at different points in the pipelin…

Read more

The hard part is everything between model.predic…

Read more