alpha-lens

The autopsy your backtest deserves before you deploy.

A Python library for post-mortem analysis and production-readiness validation of quantitative trading strategies.

---

Why alpha-lens exists

A 3.0 Sharpe ratio on a five-year backtest looks compelling. It is also the single most common artifact of overfitting in quantitative research. Existing tools like alphalens and Qlib report what the backtest shows. They do not answer the question that actually matters before a strategy goes live:

Will this Sharpe ratio survive contact with reality?

alpha-lens is built around that question. Given the daily returns of a strategy — and, optionally, factors, benchmarks, positions, and the other strategies you tried before this one — it produces a single Production Readiness Score from 0 to 100 with a READY / CONDITIONAL / NOT_READY / REJECT verdict.

The score aggregates seven components:

Component	What it measures	Method
Overfitting Risk	Probability the Sharpe is selection-bias luck	Deflated Sharpe (Bailey & López de Prado 2014), PBO via CSCV, Minimum Backtest Length
OOS Validation	Does the strategy work on data it was not built on?	70/30 chronological split + 5-window walk-forward
Robustness	Is the result fragile?	1000-sample bootstrap CI on Sharpe + 4-subsample stability
Regime Robustness	Does it work in bull, bear, and high-vol regimes?	Rule-based or HMM regime detection, per-regime Sharpe
Alpha Decay	How fast does the signal die?	Rank IC across forward horizons, exponential half-life fit
Cost Sensitivity	What does it take to kill the strategy?	Bisection on transaction cost, break-even bps
Performance	Raw return profile	Sharpe, Sortino, Calmar, drawdown statistics

The weights are calibrated so the score rewards evidence the strategy will work in production, not raw backtest performance. A 3.0-Sharpe backtest with a 70% PBO and 80% OOS degradation scores in the 20s. A 1.2-Sharpe strategy with statistically significant Deflated Sharpe, low PBO, stable walk-forward, and a 50bps cost margin scores in the 80s.

---

Quick start

pip install alpha-lens

import pandas as pd
from alpha_lens import autopsy

# Daily returns of your strategy.
returns = pd.Series(...)

report = autopsy(returns)
print(report.readiness.verdict)         # ReadinessVerdict.CONDITIONAL
print(report.readiness.overall_score)   # 68.3
report.save("autopsy.html")             # Standalone interactive HTML.

That's it. The report opens in any browser, works offline (Plotly.js is inlined), and contains six interactive tabs of diagnostics.

For a richer autopsy, pass everything you have:

report = autopsy(
    returns,
    benchmark_returns=spy_returns,    # Enables CAPM-style attribution.
    factors=ff3_factors,              # Multi-factor regression.
    factor_values=my_signal,          # IC and alpha decay analysis.
    positions=daily_weights,          # Accurate cost analysis.
    strategy_variants=other_variants, # PBO via CSCV.
    n_trials_assumed=500,             # How many things did you actually try?
)

---

Inside the report

The HTML report has six tabs. Each is designed to be read in <30 seconds.

Overview — the headline

Equity curve, rolling 63-day Sharpe with regime overlay, and a calendar heatmap of monthly returns. The first thing a PM looks at.

Production Readiness — the differentiator

Where alpha-lens earns its name. Every diagnostic that goes into the score, shown with the value, the threshold, and a plain-English interpretation. Deflated Sharpe, PBO, minimum backtest length, walk-forward Sharpes, bootstrap CI, cost-breakeven sweep — the whole audit trail.

Regime, Drawdowns, Attribution, Decay

Four drill-down tabs. The drawdown tab tags each peak-to-trough event with the dominant regime that caused it (weighted by loss magnitude, not just frequency). The decay tab fits an exponential half-life to the IC term structure. Attribution runs an OLS regression with t-stats and reports an uniqueness score.

📊 Live demo: open docs/demo/sample_report.html in any browser to see the full interactive report.

---

How is this different from alphalens and Qlib?

	alphalens	Qlib	alpha-lens
Factor IC + decay	✅	✅	✅
Returns by quantile	✅	✅	(not the focus)
Turnover and cost analysis	partial	✅	✅ + bisection for break-even
Regime decomposition	❌	❌	✅ rule-based + HMM
Deflated Sharpe Ratio	❌	❌	✅
PBO via CSCV	❌	❌	✅
Minimum Backtest Length	❌	❌	✅
Walk-forward consistency	❌	partial	✅
Bootstrap Sharpe CI	❌	❌	✅
Production Readiness Score	❌	❌	✅
Standalone interactive HTML report	partial	❌	✅ <6MB, offline-capable

alpha-lens is not a replacement for alphalens. It complements it: alphalens is best when you are researching a factor; alpha-lens is best when you are deciding whether to trade it.

---

Methodology

The statistical machinery is grounded in published research:

• Deflated Sharpe Ratio — Bailey, D. and López de Prado, M. (2014). The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality. Journal of Portfolio Management.
• Probability of Backtest Overfitting — Bailey, D., Borwein, J., López de Prado, M., and Zhu, Q. (2017). The Probability of Backtest Overfitting. Journal of Computational Finance.
• Minimum Backtest Length — Bailey, D. and López de Prado, M. (2014). The Sharpe ratio efficient frontier. Journal of Risk.
• Information Coefficient — Grinold, R. and Kahn, R. (1999). Active Portfolio Management. (Rank-IC via Spearman.)
• Walk-forward analysis — common in quant practice; see Pardo (2008).

Implementation notes are in docs/concepts.md.

---

Examples

The examples/ directory contains:

• quickstart.py — 20-line autopsy on a synthetic strategy.
• momentum_autopsy.py — full cross-sectional momentum L/S autopsy with benchmark, factors, factor values, and positions.
• compare_real_vs_lucky.py — the canonical "best-of-N noise" experiment. Shows how PBO and Deflated Sharpe expose lucky backtests.

Run any of them:

python examples/quickstart.py
open out/quickstart.html

---

Installation

# Core library.
pip install alpha-lens

# With optional HMM-based regime detection.
pip install "alpha-lens[hmm]"

# Development install.
git clone https://github.com/alpha-lens/alpha-lens
cd alpha-lens
pip install -e ".[dev]"
pytest

Python 3.10+. Core deps: numpy, pandas, scipy, scikit-learn, plotly, statsmodels, pydantic.

---

Project status

Version 0.1.0 — usable, tested, and documented. 85 tests covering every analysis module. The API is not yet frozen; expect refinements in the scoring weights and report layout in the next few releases.

Roadmap:

• Capacity / market-impact estimation (currently we report turnover only)
• Optional LLM-based interpretation layer (translate the scorecard into prose)
• Built-in adapters for Qlib and vectorbt output formats
• Stationary bootstrap option for time-series CI

---

License

MIT. See LICENSE.

Acknowledgements

The Production Readiness Score is heavily influenced by Marcos López de Prado's work on backtest overfitting. Any errors or oversimplifications in the implementation are mine.