AI Quant Trading Platform — Phase 1

A backtesting research platform built around one central principle: a backtest is only as trustworthy as its leakage-safety and its honesty about what its own numbers mean. Phase 1 covers the data layer, technical indicators, two trading strategies, a bar-by-bar backtesting engine with walk-forward validation, and an evaluation run against 10 years of real market data.

This README reports what was actually found — including a strategy that looked strong on paper but wasn't, and a methodological trap that would have gone unnoticed without an explicit diagnostic built to catch it.

What's implemented

• Data layer (src/data_layer/): fetches and locally caches daily OHLCV data (via yfinance), with an NYSE trading-calendar-aware gap detector and a non-destructive data-quality report (flags zero-volume rows, NaN OHLC values, and unexplained calendar gaps without silently dropping them).
• Indicators (src/indicators/): EMA, RSI, MACD, Bollinger Bands, ATR — hand-implemented rather than pulled from a library, with RSI and ATR using true Wilder's smoothing (simple-average seed, then recursive smoothing).
• Strategies (src/strategies/): an EMA-crossover momentum strategy and a Bollinger/RSI mean-reversion strategy, both built against a Strategy interface whose leakage-safety is verified by tests that assert a signal never changes when the strategy is truncated to an earlier subset of data.
• Backtest engine (src/backtest/): bar-by-bar simulation with next-bar-open execution (a signal generated from data through bar t is never filled at bar t's own close — only at bar t+1's open), ATR-based fixed-fractional position sizing, and a parameterized transaction-cost model.
• Walk-forward validation (src/backtest/walk_forward.py): rolling, strictly non-overlapping train/test windows; only test-window performance is ever reported.
• Metrics (src/backtest/metrics.py): Sharpe, Sortino, CAGR, max drawdown, win rate, profit factor — each with explicit, documented behavior for edge cases (empty trade lists, zero-volatility curves, all-losing trades) rather than silent NaN propagation or crashes.

Methodology

• Data: 10 years of daily bars for SPY, AAPL, and BTC-USD, fetched directly from Yahoo Finance.
• Cleaning: rows flagged as having NaN OHLC values are dropped before backtesting (an undefined price can't produce a defined fill). Zero-volume rows are intentionally kept — zero volume isn't automatically bad data. Calendar gaps are left untouched, since they're simply dates absent from the source, not bad rows to fix.
• Walk-forward windows: 2-year training period, 6-month non-overlapping test period, over the full 10-year history — 17 windows per asset. Only test-window results are ever used for reported metrics; training windows never contribute to a reported number.
• Costs: every combination was run twice — once at a 10 bps baseline transaction cost, once at a 25 bps stress scenario — specifically to surface how much of any apparent edge survives realistic friction.
• Position sizing: fixed-fractional, sized so that an ATR-based stop distance risks 1% of current equity per trade.
• Linking across windows: each window's equity curve is independently simulated starting from a fixed initial capital, then scaled and chained onto the previous window's ending equity — this preserves each window's true percentage return while avoiding a misleading capital reset at every 6-month boundary.

The main finding: a strategy's Sharpe ratio can be almost entirely fake

The backtest engine leaves a position open (unrealized) if it's still active at the last bar of a test window, rather than forcing an artificial close. This is the right behavior for realism — but it creates a subtle trap: equity-curve-based metrics (Sharpe, Sortino, CAGR) mark unrealized positions to market, while trade-level metrics (win rate, profit factor) only count trades that have actually closed. A strategy can show a strong equity curve almost entirely because of favorable unrealized positions at window boundaries, while every trade it has actually closed lost money.

To catch this, an unrealized-P&L reconciliation diagnostic was added: for each window, the equity change unexplained by realized (closed) trade P&L is attributed to open positions, and rolled up across the full evaluation.

This caught something real. Across all three assets, the EMA-crossover momentum strategy's closed-trade statistics tell a consistently different story than its equity curve:

Ticker	Sharpe (baseline)	Win rate	Profit factor
SPY	0.10	25.0%	0.37
AAPL	0.40	15.9%	0.30
BTC-USD	1.10	25.4%	0.71

Every single momentum result has a win rate under 26% and a profit factor under 1 — meaning realized losses exceed realized gains on every closed trade, on every asset, under every cost scenario tested. The positive Sharpe ratios above are driven substantially or (for BTC-USD, where the unrealized share of total gain exceeds 100%) entirely by unrealized open-position marks at window boundaries, not by a demonstrated trading edge. Without the reconciliation diagnostic, this would have read as three solid, profitable results.

Honest results by asset and strategy

(baseline = 10 bps transaction cost, stress = 25 bps)

Ticker	Strategy	Cost	Sharpe	Win rate	Profit factor	Verdict
SPY	Momentum	baseline	0.10	25.0%	0.37	No real edge — see above
SPY	Momentum	stress	-0.02	22.7%	0.30	Negative under cost
SPY	Mean reversion	baseline	-0.03	31.6%	0.91	Unprofitable
SPY	Mean reversion	stress	-0.42	26.3%	0.43	Unprofitable
AAPL	Momentum	baseline	0.40	15.9%	0.30	No real edge — see above
AAPL	Momentum	stress	0.33	15.9%	0.27	No real edge — see above
AAPL	Mean reversion	baseline	0.29	57.8%	1.53	Best result — real, if thin, edge
AAPL	Mean reversion	stress	0.01	53.3%	1.06	Edge nearly erased by cost
BTC-USD	Momentum	baseline	1.10	25.4%	0.71	No real edge — see above
BTC-USD	Momentum	stress	1.06	23.9%	0.67	No real edge — see above
BTC-USD	Mean reversion	baseline	-0.75	51.6%	0.46	Unprofitable
BTC-USD	Mean reversion	stress	-0.86	43.6%	0.40	Unprofitable

The one genuinely defensible result: AAPL mean reversion (Bollinger Bands + RSI) is profitable at realistic cost (profit factor 1.53, win rate 57.8%) and, unlike the momentum results, its edge is backed by real closed trades rather than unrealized marks. It's a thin edge — it nearly vanishes under the 25 bps stress scenario — but it's real. Every other strategy/asset combination tested here is either unprofitable outright or depends on an accounting artifact rather than a demonstrated trading edge.

This is reported as the actual finding of Phase 1, not softened: the platform did not discover a robust, general trading edge. What it did demonstrate is a rigorous, leakage-safe evaluation pipeline capable of telling the difference between a real edge and a mirage — which is the harder and more valuable engineering problem.

Known limitations

• Indicator cold-start per window: each walk-forward test window computes its indicators independently, with no carried-over history from the training window. This forces roughly the first 20-26 rows of every window to be flat while indicators warm up — real evaluation time lost across 17 windows per asset. A more complete version would extend each window's data backward by the indicator lookback and only start counting trades/equity from the true test-window start.
• Exit-side slippage uses entry-time ATR, not ATR at the moment of exit. If volatility regime changes meaningfully between entry and exit, this under- or overstates exit slippage. A deliberate simplification, not an oversight — worth revisiting if slippage sensitivity becomes a focus.
• The unrealized-P&L reconciliation is an approximation, not exact accounting — it nets out costs already embedded in closed-trade P&L and attributes whatever equity change is left over to open positions. It becomes numerically unstable (large or meaningless ratios) when a window's total gain is near zero, so it's most informative for large, unambiguous gains and should be read alongside win rate/profit factor rather than in isolation.
• Only two strategies and three assets were tested. These results characterize this specific strategy design and cost model, not momentum or mean-reversion trading in general.

What's next (Phase 2+)

• ML-based signal generation (XGBoost/LightGBM) as an alternative to hand-coded indicator rules
• Market-regime detection and strategy selection
• A proper risk-management layer (dynamic stops, portfolio-level diversification, max-drawdown circuit breakers)
• Paper trading loop
• Dashboard with live signals and the AI-explanation layer

Reproducing these results

python scripts/run_phase1_evaluation.py

Full, unrounded results — including the per-window breakdown behind every aggregate number in the tables above — are saved to results/phase1_evaluation_{date}.json.