A self-contained research harness that implements 50 trading signals, each traced to a specific academic paper, and evaluates them with a single rigorous walk-forward backtest across equities, FX, commodities, crypto and volatility products.
Built for QuantiHack 2026 (theme: data manipulation) — the goal was to turn unconventional, "alternative" data sources into testable signals and rank them honestly, fast, under a hackathon deadline.
Most strategy backtests cherry-pick one signal and one period. This does the opposite:
-
50 signals, one framework. Every strategy is a small
signal()function plugged into a shared backtest engine (position sizing, slippage, equity curve, metrics) so they're compared on identical footing. -
Every signal is sourced. Each one cites the paper it comes from — Tetlock (2007) on media pessimism, Ariel (1990) on the holiday effect, Asness et al. (2012) on risk parity, Amihud (2002) on illiquidity, and 46 more — grouped into six families:
Family Examples Textual / NLP media-pessimism reversal, Wikipedia-attention proxy Calendar & temporal holiday effect, Monday reversal, quarter-end window dressing Cross-asset leakage ETF-flow reversal, gold crisis hedge Microstructure & order flow odd-lot retail contrarian, Amihud illiquidity premium Behavioural & sentiment round-number avoidance, VIX-percentile contrarian Volatility surface implied–realised vol spread, kurtosis tail premium -
Two-window walk-forward. Every signal is scored on the full history (WF1: 2010 → 2026) and re-run on a deliberate out-of-sample stress window (WF2: Feb–Apr 2026, the tariff-war drawdown) to surface what survives a regime it was never tuned on.
Signals are ranked by combined = avg(WF1 Sharpe rank, WF2 return rank) — rewarding
strategies that are both strong in-sample and robust through the crisis window, not just
one or the other. Full metrics per strategy (CAGR, Sharpe, Sortino, max drawdown, Calmar,
profit factor, skew, beta, alpha) are written to quantihack_alt_data_50_results.csv.
| Signal | Source | WF1 Sharpe | WF2 return |
|---|---|---|---|
| Conditional Risk Parity | Asness et al. (2012) | 0.76 | +1.96% |
| Parkinson Regime Switch | Parkinson (1980) | 0.64 | +4.37% |
| Holiday Effect | Ariel (1990); Lakonishok & Smidt (1988) | 0.60 | +4.80% |
| Round-Number Avoidance | Bhattacharya et al. (2012) | 0.69 | +1.10% |
| Volume-Weighted RSI(2) | Lerman et al. (2008) | 0.61 | +2.46% |
pip install -r requirements.txt
python quantihack_alt_data_50.py # full backtest → results CSV + chart PNGquantihack_algo_baseline.py is a separate multi-asset baseline bot (16 symbols, position
limits, stop-loss / take-profit / cooldown) used as a sanity reference.
Python 3 · pandas / numpy / scipy · yfinance · matplotlib
- Research code, not a live trading system — no broker integration, fills are modelled with a flat per-side slippage.
- Data is pulled via
yfinance(with local CSV fallbacks for SPY/TLT/VIX/GLD/BTC/ETH). - The point is breadth + honest validation: 50 ideas, one engine, an out-of-sample stress test, and a leaderboard you can argue with.