Signals

Quant research platform for US equity momentum. The core finding: expanding the stock universe from S&P 500 to S&P 500+400 is the single most impactful improvement — lifting Sharpe from 0.659 to 1.075 and CAGR from 13.6% to 28.4%. This survives deflated Sharpe correction at n_trials=3 (DSR=1.000) and four independent survivorship bias validations.

Forward evidence (2026-04-24): 10 trading days of live paper trading on Alpaca. Momentum account +4.79% vs SPY +4.27% (n=10, not statistically significant). See LIVE_RECORD.md and SKEPTIC_REVIEW_V2.md for why the backtest's headline numbers are probably inflated.

Disclaimer

Experimental research. Not financial advice. Backtest results are historical and do not predict future performance. The early-period (2000-2010) results show a survivorship bias signature (Test 1 below). The recommended production config limits the universe to former S&P 500 constituents, which reduces this bias. MIT-licensed, no warranty. See LICENSE.

---

Results

26-year backtest (2000-2026) using total-return (split + dividend adjusted) prices. Train: 2000-2018. Holdout: 2019-2026.

Performance

Strategy	Sharpe	CAGR	Max DD	Calmar	Holdout Sharpe
Full-universe momentum (S&P 500+400)	1.075	28.4%	-55.2%	0.51	0.708
Former-S&P-only (conservative)	0.998	26.4%	-57.6%	0.46	—
S&P-only (point-in-time constituents)	0.659	13.6%	-62.9%	0.22	0.530
SPY buy & hold (total return)	0.497	8.0%	-55.2%	0.14	—

Statistical significance (Deflated Sharpe Ratio)

n_trials	DSR	Verdict
1	1.000	Pass
2	1.000	Pass (honest trial count for universe choice)
3	1.000	Pass
4	0.833	Fail

The full-universe variant is a single structural choice (constituent filter: yes/no), not a parameter sweep. DSR@2 is the honest test.

Survivorship bias validation

Test	Result	Details
Time-decay	CONCERN	Full-universe advantage shrinks from +0.46 (2000-05) to -0.05 (2020-26). Early years likely inflated.
Delisting simulation	PASS	Killing 3%/year with -50% terminal returns: Sharpe 0.888 +/- 0.063. Worst seed (0.788) still beats S&P-only (0.659).
Former-S&P-only	PASS	112% of improvement from stocks that were in S&P historically. Never-S&P stocks are poor (0.488).
Sector concentration	PASS	HHI 0.090 (near-uniform across 12 sectors). No sector loading.

The recommended production config is former-S&P-only (Sharpe 0.998) — gets most of the full-universe benefit while limiting survivorship exposure to stocks that were actually in a major index.

See scripts/survivorship_validation.py for the full validation suite.

---

How the model works

Instead of buying stocks with the highest trailing 12-month return (classic momentum — which buys at the top), the early-breakout model ranks by momentum acceleration: 3-month return minus the annualized 12-month pace. This catches stocks at the start of a move.

• Signal: 3-month/12-month acceleration (63d/252d windows)
• Filter: min 10% short-window return, max 150% long-window return
• Universe: all stocks with price data (S&P 500+400, ~1,100 tickers)
• Diversification: max 2 per GICS sector, 15 positions
• Hold: 105 trading days (~5 months), fixed
• Fully invested: 0% cash reserve, contributions deployed immediately
• Costs: 10 bps round-trip (5 bps commission + 5 bps slippage)

Universe scaling (the key finding)

Universe	Tickers	Sharpe	CAGR	Holdout Sharpe
S&P 500 only	~500	0.659	13.6%	0.530
S&P 500+400	~1,100	1.075	28.4%	0.708
S&P 500+400+600	~1,400	1.072	28.5%	0.739
Broad US (4,000+)	~2,200	0.986	25.7%	0.738

The S&P 500+400 universe is the sweet spot. Adding micro-caps beyond that dilutes selection quality. More stocks gives the acceleration signal a richer tail to select from — consistent with academic evidence that momentum is stronger in mid-caps (less analyst coverage, slower information diffusion).

What we tested and rejected

Approach	Result
Quality factors (GP/A, low-vol, ROE)	Penalize the breakouts the strategy targets
Risk-managed sizing (Barroso/Santa-Clara)	Fights the fixed-hold exit mechanism
Residual momentum (beta-adjusted)	Existing filters already capture stock-specific signal
Value factor (mean-reversion tilt)	Anti-momentum, dilutes the signal
Classic 12-month momentum	Works but survivorship bias inflates results ~80%
26 exit rules (profit targets, stop losses, trailing stops)	None beat doing nothing
Regime filters (golden cross, drawdown limits)	Reduce drawdown but cost too much CAGR
Markov chains (5 variants, 100+ tests on BTC)	Tied with buy & hold (Sharpe ~1.0)
Trend filters / golden cross on individual stocks	0/20 beat buy & hold
Pairs trading (stat arb)	Negative returns after costs
Portfolio optimizers (risk parity, mean-variance)	No improvement over equal weight
108-config parameter sweep	All within noise of canonical config

BTC Markov chain (exhausted track)

The project started as a BTC Markov-chain signal generator. After extensive testing (Tier 0-3, 1,600+ backtests, 5 model classes), the BTC hybrid model achieves multi-seed average Sharpe ~1.0 (tied with buy & hold) with lower drawdowns. A null hypothesis test confirmed the Markov chain adds +0.37 Sharpe over a naive vol filter — it's not decoration, but it's a risk management tool, not alpha.

See scripts/NULL_HYPOTHESIS_RESULTS.md and SKEPTIC_REVIEW.md.

---

Install

# From GitHub
pip install git+https://github.com/jlgreen11/signals.git

# Or clone for development
git clone https://github.com/jlgreen11/signals.git
cd signals
python3.12 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Data setup

# Download S&P 500+400 price data (~15 min first time)
python scripts/download_sp400.py

# Optional: S&P 600 SmallCap
python scripts/download_sp600.py

API keys (optional, for paper trading)

Create a .env file (gitignored):

ALPACA_API_KEY=PK...
ALPACA_SECRET_KEY=...
ALPACA_BASE_URL=https://paper-api.alpaca.markets

Get keys at alpaca.markets → Paper Trading → API Keys.

Usage

Backtest

from signals.backtest.bias_free import load_bias_free_data, run_bias_free_backtest

data = load_bias_free_data()
result = run_bias_free_backtest(data, use_full_universe=True)

print(f"Sharpe: {result.sharpe:.3f}")
print(f"CAGR:   {result.cagr:.1%}")
print(f"MaxDD:  {result.max_drawdown:.1%}")

Signal generation

from signals.model.momentum import CrossSectionalMomentum

mom = CrossSectionalMomentum(
    mode="early_breakout", lookback_days=252,
    short_lookback=63, n_long=15, max_per_sector=2,
)
weights = mom.rank(prices_dict, as_of_date=pd.Timestamp("2026-04-21", tz="UTC"),
                   sectors=sector_map)

Automated trading (Alpaca)

signals auto daily --account momentum    # generate signals
signals auto trade --account momentum    # execute paper trades
signals auto performance --account all   # compare accounts
signals auto positions --account momentum

Full-universe auto-discovery: the automation layer scans data/raw/ for all available tickers. Supports multiple parallel accounts with monthly rebalancing tracked in SQLite.

Tests

pytest --cov=signals    # 342 tests, ~50 seconds
ruff check signals tests

Key files

File	Purpose
signals/backtest/bias_free.py	Canonical backtest engine (single source of truth)
signals/model/momentum.py	Cross-sectional momentum (classic + early-breakout)
signals/automation/	Paper trading, signal blending, daily execution
scripts/survivorship_validation.py	4-test survivorship bias validation suite
scripts/FULL_UNIVERSE_RESULTS.md	Full-universe evaluation results
scripts/FULL_EVALUATION_2026_04_21.md	Comprehensive evaluation write-up
SKEPTIC_REVIEW.md	Methodology critique and red flags
COMPREHENSIVE_EVALUATION.md	108-config grid sweep results
FUTURE_IMPROVEMENTS.md	Tested and pending ideas

History

Started April 2026 as a Markov-chain BTC signal generator. A skeptic review exposed inflated results. Pivoted to cross-sectional equity momentum. Multiple iterations corrected survivorship bias, switched to total-return prices, and established a canonical backtest module.

Key milestones:

• Apr 10-11: BTC Markov chain exhausted (Sharpe plateau at ~1.0)
• Apr 16: 108-config grid sweep; DSR fails across all configs
• Apr 21: Full-universe breakthrough (Sharpe 0.659 → 1.075); null hypothesis test confirms BTC Markov chain is real but marginal; survivorship bias validation (3/4 pass); S&P 400+600 expansion; 342 tests passing