Residual Edge Analyzer

Overview

Test whether a strategy's apparent performance survives explicit comparison with predeclared baseline return series. Produce an auditable JSON artifact and a concise Markdown report without fetching data or changing trading exposure.

Treat this as a falsification gate after backtest-expert, not as trade authorization.

Prerequisites

• Use Python 3.9+.
• Prepare one CSV containing an ISO date, strategy return, and every baseline return on

the same row.

• Prepare a JSON specification following

[references/input-contract.md](references/input-contract.md).

• Supply actual period returns. Do not substitute CAGR, Sharpe, cumulative P&L, or other

summary metrics.

Workflow

1. Define the question before inspecting results

State the claimed independent edge in one sentence. Select a primary baseline that is a plausible simple copy of the strategy, then select at least one alternate baseline model.

Record these declarations in the config:

• baseline_selection: predeclared
• strategyreturnbasis and baselinereturnbasis: both gross or both net
• analysisscope: outofsample, live, or insample
• universedata: pointintime, currentconstituents, or not_applicable

Every declaration is mandatory for a decision-grade verdict. Omitting one is treated as undeclared, not as benign, and drops the report to REVIEWREQUIRED. notapplicable exists so that a baseline with no universe membership can be declared explicitly rather than left blank.

Do not choose a baseline because it gives the preferred residual result.

2. Validate the return-series contract

Require:

• unique ISO dates;
• finite numeric returns greater than -100%;
• identical frequency and cost basis across strategy and baselines;
• point-in-time membership for same-universe equal-weight or momentum baselines;
• regime labels defined independently of the loss periods being explained.

Stop if the input lacks a dated strategy return series. Report summary-only input as insufficient rather than inventing observations.

3. Run the analyzer

python3 skills/residual-edge-analyzer/scripts/analyze_residual_edge.py \
  --input reports/strategy_returns.csv \
  --config reports/residual_edge_config.json \
  --output-json reports/residual_edge_report.json \
  --output-markdown reports/residual_edge_report.md

The script runs the predeclared primary model and all sensitivity models in one execution. It uses an intercept OLS model and HAC/Newey-West standard errors. It reports the residual edge ratio as annualized alpha divided by annualized residual volatility; do not calculate a Sharpe ratio from raw OLS residual mean because an intercept makes that mean zero.

4. Interpret the evidence

Use the four statuses as diagnostic labels:

• RESIDUAL_EDGE: alpha, residual edge ratio, and rolling stability clear configured

thresholds.

• BASELINE_EXPLAINED: baseline R-squared is high while residual evidence is weak.
• RESIDUAL_FRAGILE: results fail one or more robustness gates or change across declared

baseline models. Also use this status when rolling analysis is disabled, unavailable, incomplete, or no sensitivity model was supplied.

• INSUFFICIENT_EVIDENCE: the sample is below the configured minimum.

Read decisioneligibility separately. A statistically interesting result remains REVIEWREQUIRED when critical provenance, cost-basis, sample, or multicollinearity warnings exist, when rolling evidence is unavailable, or when no alternate baseline was tested.

Inspect:

1. primary and sensitivity-model status;
2. annualized alpha and HAC t-stat;
3. residual edge ratio and residual autocorrelation;
4. rolling alpha stability;
5. VIF for multi-factor models;
6. active-return breakdown across predeclared regimes.

5. Hand off findings

• Send baseline-choice, OOS, and stability findings back to backtest-expert.
• Send recurring residual failure regimes to signal-postmortem.
• Pass only evidence and operating constraints to trade-performance-coach.
• Never change position size, exposure, or orders automatically.

Boundaries

• Do not call this holdings-based contribution analysis. Brinson allocation, selection,

and interaction effects require historical holdings, benchmark weights, and constituent returns.

• Do not claim stock-selection alpha from a market-index-only baseline.
• Do not build equal-weight baselines from current constituents and label them

point-in-time.

• Do not interpret in-sample residual edge as confirmed alpha.
• Do not mine many regime definitions after seeing losses. Predeclare a small set and

confirm findings out of sample.

• Do not assume high R-squared makes a strategy worthless; capacity, tail behavior, costs,

and implementation value require separate evidence.

Resources

• scripts/analyzeresidualedge.py — deterministic CSV-to-JSON/Markdown analyzer.
• references/input-contract.md — CSV/config contract and runnable example.
• references/methodology.md — statistical definitions, interpretation, and limitations.