Bias Taxonomy

Bias Taxonomy turns the bias taxonomy — the checklist of errors that inflate backtests — into running experiments, the first chart-native implementation of the concept. A fixed SMA 20/50 reference strategy is put through paired tests: fills at the signal close versus shifted one bar (look-ahead), pivot signals taken as redrawn versus once confirmed (repainting), the best sweep variant versus the median (data snooping), and returns gross versus net of a per-side cost (cost omission). Survivorship is listed as MANUAL — not measurable from one surviving symbol. Each row shows both legs and a severity status.

How to Trade the Bias Taxonomy?

• HIGH status: the audited leg's gain disappears entirely — the inflated result was bias, not edge.
• CAUTION status: the audited leg gives up at least the severity-threshold share of the inflated gain.
• Bias Inflation Wedge: the fill between naive equity (best variant, leaky fills, gross) and audited equity (pre-registered, shifted, net) — accumulated upward bias made visible.
• OMITTED on the cost row: the cost input is zero, so the audit's own cost entry is still open.

Bias Taxonomy Settings

• Fast MA Length (default 20) and Slow MA Length (default 50): the pre-registered reference strategy.
• Per-Side Trading Cost (%) (default 0.05): charged on every position change.
• Repaint Audit Pivot Length (default 5): swing width of the repaint audit; the confirmed leg waits this many bars.
• Snooping Sweep Fast: Start (default 5), Step (default 5), Count (default 4), and Snooping Sweep Slow: Start (default 20), Step (default 20), Count (default 4): the grid of tested variants.
• Severity Threshold (%) (default 50): inflated-gain share surrendered before a row turns CAUTION.
• Show Bias Audit Table (default enabled) with position/size, and Show Equity Curves (default enabled).

Frequently Asked Questions

What should I do when the data-snooping row flags?

Treat the winner as manufactured until proven otherwise: cut and pre-register trials, then validate on data the selection never saw — the rolling out-of-sample discipline Walk-Forward Analysis implements.

Why is survivorship always MANUAL?

A universe defined by today's membership silently excludes the delisted graveyard, and nothing computed on one surviving symbol can detect that. The row stays so the audit is honest about its blind spot.

Does this audit my own strategy?

It audits the fixed reference strategy — deliberately. The measured inflation shows how strongly each bias distorts results on this symbol and timeframe, calibrating skepticism for backtests built on the same data.