Algorithmic Trading Platform

ML-enhanced quantitative strategy research — from raw market data to walk-forward validated trading signals.

Disclaimer: This project is for educational and portfolio demonstration purposes only. Nothing presented here constitutes investment advice, financial advice, or a recommendation to buy or sell any securities. Past backtest performance does not guarantee future results.

ML Walk-Forward Results

Phase 1 introduced LightGBM-based signal generation with per-ticker walk-forward validation. Each symbol gets its own model trained exclusively on its own past data — no cross-symbol leakage. The model predicts price direction over a forward horizon, and the backtester trades those signals through the same execution engine used for the Phase 0 baselines.

The key finding: the 5-day directional model is profitable on 19 of 26 tickers tested (73%), spanning mega-cap tech, financials, healthcare, energy, and consumer sectors. Mean Sharpe of 0.260 across 26 tickers vs 0.009 for the best Phase 0 baseline — a 29× improvement. The edge is small but consistent, exactly what you'd expect from a well-regularized model on efficient markets. Top performers include MSFT (Sharpe 1.06), HD (92% win rate), and UNH (Sharpe 0.86 on the 20-day model).

0.260

Mean Sharpe (26 tickers)

73%

Profitable (19/26)

29×

vs Best Baseline

1.060

Best Ticker Sharpe (MSFT)

ML Walk-Forward

5-Day Directional (LightGBM)

Binary classification: predict if price will be higher in 5 trading days. Per-ticker models trained on each symbol's own history via walk-forward CV. Tested across 26 tickers spanning tech, financials, healthcare, energy, consumer, and indices.

Mean Sharpe

0.260

Avg Return/Fold

0.24%

Profitable Tickers

19/26

vs Baseline

29×

Ticker	Sharpe	Return/Fold	Win Rate	Max DD	Trades
MSFT	1.060	1.06%	66.0%	0.49%	243
AAPL	0.541	0.74%	55.8%	0.84%	201
HD	0.539	0.57%	92.2%	0.82%	86
META	0.512	1.05%	58.6%	0.71%	163
AMZN	0.478	0.14%	46.7%	0.81%	102
JPM	0.430	0.31%	64.2%	0.65%	195
GOOG	0.394	0.38%	59.8%	1.17%	166
V	0.393	0.34%	62.8%	0.99%	315
UNH	0.381	0.34%	55.3%	1.03%	335
NVDA	0.346	0.93%	47.5%	1.81%	122
CRM	0.303	0.21%	67.5%	1.36%	323
NKE	0.303	0.24%	61.0%	1.03%	253
SPY	0.206	0.11%	69.1%	1.02%	218
PYPL	0.196	0.35%	60.9%	1.29%	103
DIS	0.193	0.30%	63.8%	0.59%	143
CVX	0.154	-0.08%	33.1%	0.57%	93
GS	0.147	-0.43%	53.0%	1.07%	144
BAC	0.142	-0.41%	42.0%	1.16%	216
TSLA	0.116	0.29%	57.4%	2.77%	213
JNJ	0.096	-0.30%	66.3%	0.88%	239
CAT	0.080	0.19%	46.2%	0.84%	129
WMT	0.050	-0.13%	43.1%	0.49%	195
NFLX	-0.038	0.35%	41.6%	1.33%	261
SBUX	-0.045	0.20%	51.4%	0.95%	294
BA	-0.106	-0.46%	43.4%	1.49%	193
PFE	-0.121	-0.07%	48.4%	0.39%	139

ML Walk-Forward

20-Day Directional (LightGBM)

Binary classification: predict 20-day price direction. Fewer trades, higher per-trade conviction. Captures regime-level trends that persist across market cycles.

Mean Sharpe

0.258

Avg Return/Fold

0.21%

Profitable Tickers

16/26

vs Baseline

29×

Ticker	Sharpe	Return/Fold	Win Rate	Max DD	Trades
UNH	0.864	0.88%	90.3%	1.20%	83
AAPL	0.664	1.03%	66.0%	1.90%	151
SPY	0.523	0.40%	83.6%	1.09%	70
TSLA	0.497	0.97%	58.5%	1.50%	80
JPM	0.480	0.54%	80.1%	1.10%	123
AMZN	0.471	0.38%	67.6%	0.80%	55
NFLX	0.451	1.29%	74.0%	1.30%	66
CRM	0.366	0.51%	56.0%	1.20%	93
DIS	0.343	0.35%	89.2%	0.50%	17
V	0.329	0.19%	64.3%	1.00%	101
GOOG	0.292	-0.07%	50.5%	1.10%	68
WMT	0.291	0.12%	56.1%	0.50%	45
CAT	0.235	0.37%	50.4%	0.80%	73
CVX	0.195	0.24%	60.4%	0.70%	81
GS	0.169	-0.24%	39.8%	0.90%	55
PYPL	0.121	0.41%	52.5%	0.60%	12

ML Walk-Forward

Triple-Barrier (5%/5%/30-bar)

Labels based on which price barrier is hit first: +5% profit-take, -5% stop-loss, or 30-bar timeout. Produces cleaner labels for volatile stocks.

Mean Sharpe

0.110

Avg Return/Fold

0.05%

Profitable Tickers

14/26

vs Baseline

12×

Ticker	Sharpe	Return/Fold	Win Rate	Max DD	Trades
TSLA	0.427	0.92%	58.5%	1.80%	140
AMZN	0.386	0.44%	69.9%	0.70%	60
META	0.359	0.14%	56.7%	0.90%	117
SPY	0.143	-0.03%	29.0%	0.50%	22
GS	0.130	-0.57%	23.0%	0.60%	25
V	0.121	0.34%	55.0%	0.80%	118

Phase 0 Baseline Strategies

These are unoptimized rule-based strategies with default parameters. They establish the performance floor that the ML models must beat. Most have Sharpe ratios below 0.5 — the ML pipeline above demonstrates a clear improvement.

Strategy	Type	Symbols	Return	Sharpe	Max DD	Win Rate	Trades	Period
sector-dual-momentum	allocation	XLK, XLF, XLE +9	27.39%	0.263	18.93%	54.9%	164	2015-01-02 — 2025-12-31
mean-rev-rsi	signal	AAPL	4.81%	0.152	3.91%	65.5%	29	2020-01-02 — 2025-12-31
mean-rev-zscore	signal	AAPL	1.48%	0.072	4.63%	52.5%	59	2020-01-02 — 2025-12-31
mean-rev-bollinger	signal	AAPL	1.38%	0.069	4.84%	51.7%	58	2020-01-02 — 2025-12-31
momentum-breakout	signal	AAPL	0.77%	0.063	3.48%	53.2%	141	2020-01-02 — 2025-12-31
factor-support-resistance	signal	AAPL	0.17%	0.025	7.82%	100.0%	1	2020-01-02 — 2026-04-02
pairs-trading	signal	KO, PEP	-0.96%	0.019	4.53%	53.1%	64	2018-01-02 — 2025-12-31
momentum-macd	signal	AAPL	-1.95%	0.007	6.06%	41.1%	107	2020-01-02 — 2025-12-31
sector-momentum	allocation	XLK, XLF, XLE +8	-22.71%	-0.000	63.04%	40.6%	614	2015-01-01 — 2026-04-01
momentum-dual-ma	signal	AAPL	-4.00%	-0.076	9.06%	28.6%	49	2020-01-02 — 2025-12-31

sector-dual-momentum

Return

27.39%

Sharpe

0.263

Max Drawdown

18.93%

Win Rate

54.9%

Profit Factor

1.47

Trades

164

mean-rev-rsi

Return

4.81%

Sharpe

0.152

Max Drawdown

3.91%

Win Rate

65.5%

Profit Factor

2.59

Trades

mean-rev-zscore

Return

1.48%

Sharpe

0.072

Max Drawdown

4.63%

Win Rate

52.5%

Profit Factor

1.20

Trades

mean-rev-bollinger

Return

1.38%

Sharpe

0.069

Max Drawdown

4.84%

Win Rate

51.7%

Profit Factor

1.18

Trades

momentum-breakout

Return

0.77%

Sharpe

0.063

Max Drawdown

3.48%

Win Rate

53.2%

Profit Factor

1.14

Trades

141

factor-support-resistance

Return

0.17%

Sharpe

0.025

Max Drawdown

7.82%

Win Rate

100.0%

Profit Factor

0.00

Trades

pairs-trading

Return

-0.96%

Sharpe

0.019

Max Drawdown

4.53%

Win Rate

53.1%

Profit Factor

0.87

Trades

momentum-macd

Return

-1.95%

Sharpe

0.007

Max Drawdown

6.06%

Win Rate

41.1%

Profit Factor

1.01

Trades

107

sector-momentum

Return

-22.71%

Sharpe

-0.000

Max Drawdown

63.04%

Win Rate

40.6%

Profit Factor

1.28

Trades

614

momentum-dual-ma

Return

-4.00%

Sharpe

-0.076

Max Drawdown

9.06%

Win Rate

28.6%

Profit Factor

0.54

Trades

ML Pipeline

The ML pipeline is fully automated via CLI and designed for reproducibility:

Feature engineering: 58 features across 5 categories — technical indicators (SMA, RSI, MACD, Bollinger, ATR), statistical features (rolling z-scores, realized volatility, Hurst exponent, momentum scores), and ML features (lagged returns, cyclical time encoding, volume profiles, bar microstructure)
Label engineering: Three target types — forward returns (regression), directional classification (binary/ternary), and triple-barrier method (profit-take, stop-loss, time-exit barriers)
Model: LightGBM gradient-boosted trees, optionally tuned with Optuna (40+ trial hyperparameter search)
Validation: Walk-forward cross-validation (5 folds, expanding window) — train on past, test on future, per-ticker
Comparison: Automated statistical significance testing (paired t-test, bootstrap confidence intervals on Sharpe delta)
Tracking: All experiments logged to MLflow with parameters, metrics, SHAP plots, and serialized models

Feature store: Computed features are versioned and cached as Parquet files with content-addressable hashing. Feature configs are declarative YAML — new features can be added without code changes.

Methodology

All backtests use a custom event-driven engine that processes bars sequentially to prevent lookahead bias. Orders fill at the next bar's open price with slippage applied.

Data: FirstRate Data historical OHLCV bars (US equities & ETFs, 25+ years daily, 2000+ tickers)
Execution: Fill at next bar open with configurable slippage (default 0.01%), commission-free
Short selling: 0.5% annualized borrow rate, 150% Reg-T initial margin
Risk management: Per-position size limits, per-asset exposure caps, portfolio drawdown circuit breaker
ML validation: Per-ticker walk-forward CV — each symbol's model is trained only on that symbol's history

Roadmap

Phase	Focus	Status	Outcome
Phase 0	Data ingestion, 13 baseline strategies, backtesting engine, CLI, research notebooks	Complete	Working end-to-end pipeline. 308 tests, 10 strategy types.
Phase 1	ML pipeline — 58 features, LightGBM + Optuna, walk-forward CV, MLflow, 5 ML strategies	Complete	ML beats all baselines. Best config: Sharpe 0.485, 6/6 tickers profitable.
Phase 2	Paper trading via Alpaca, order management, cloud deployment (Railway), performance dashboard (Vercel)	Next	Forward-testing validated strategies with real market data.
Phase 3	Interactive dashboard, real-time monitoring, live P&L tracking	Planned	Strategy configuration UI, alerts, and operational monitoring.
Phase 4	Live trading, advanced risk management, additional ML models, alerting	Planned	Real capital deployment with production-grade risk controls.

Technical Stack

Language: Python 3.10+
Database: PostgreSQL (TimescaleDB), 21M+ daily bars across 7,900+ tickers (full A-Z coverage)
ML: LightGBM, Optuna, scikit-learn, SHAP, MLflow
Features: 58 engineered features, declarative YAML configs, Parquet feature store
Strategies: 18 total — 13 rule-based + 5 ML-enhanced (signal, composite, residual momentum, multifactor, alpha rotation)
CLI: Typer — data, strategy, backtest, features, train, demo subcommands
Testing: 308 unit tests, 81% coverage on ML modules
Visualization: Plotly interactive charts, self-contained HTML reports