RDI — basic training

The RDI takes a fixed book and shows what a regime detector's exposure overlay would have done to it. You shape the overlay with three levers — the score itself, where you cut it into states, and how hard you de-risk per state — and watch the counterfactual equity curve respond.

The pipeline

A detector scores the universe; thresholds turn the score into a state; a policy turns the state into gross exposure; that gross re-sizes the book's raw path into a counterfactual NAV. The three teal boxes are the levers you control — the detector's own dimensions, the thresholds, and the policy.

Anatomy of a regime

Real turbulence stress through 2019–2020. Your two cut lines (degraded 0.50, bad 0.80) split the continuous score into good / degraded / bad bands; the policy then maps those states to the gross step below. Move a cut line and the bands — and the gross — move with it.

Lever 1 · The score itself (detector ⚙)

Before you threshold anything, you shape the score. Every detector computes its continuous score from its own dimensions — a look-back above all (how much history the stress / breadth / drawdown integrates), plus its baked thresholds and quantiles. Checking a detector auto-reveals these params inline (the ⚙ gear toggles them); editing one re-runs the detector and reshapes the raw score live. The look-back is a real, connected, tunable dimension — not a hidden constant. The percentile view has its own look-back too: the Pct window, next to Scale, sets the history the rank is measured against.

Worked example · sweeping the look-back

Drawdown detector, its own baked state, default policy, sweeping only the look-back. A short 126-day window fires bad 5% of days and lands at 33×; a 504-day window fires 8% and reacts more slowly at 30× — different scores, different counterfactuals, before a single threshold is touched. Milder than the threshold or policy here, but it is the very score you go on to cut.

Lever 2 · Thresholds — a regime from the score (signal → state)

Tick author a regime from a score, choose a source detector, pick raw or percentile and a polarity (high = bad for stress, high = good for quality), then set your own degraded and bad cutoffs. The state is derived from your cuts, not the detector's baked logic, and the authored regime leads the score panel (its cut lines and bands are drawn there). It re-queries as you edit.

Worked example · sweeping the threshold

Same book, turbulence stress, degraded fixed at 0.50, sweeping the bad cutoff. A looser cut (bad≥0.60) fires 72% of days and de-risks the book down to 6×; a tighter cut (bad≥0.90) fires 15% and lands at 20×; holding through returns 66×. In a bull market more de-risking costs compounding — so read the underwater drawdown alongside CF, never CF alone.

Lever 3 · Policy (state → gross)

The Policy control sets the gross held in each state (default good 1.0 · degraded 0.65 · bad 0.15). It decides how hard a bad call actually bites. Editing it moves the gross step and every counterfactual instantly.

Worked example · tuning the policy

Same fixed cuts (0.50 / 0.80), sweeping only the bad gross. Cutting to 15% in bad lands at 14×; a gentle 50% at 20×; staying fully invested in bad at 31×. The policy is as powerful as the threshold — tune all three levers.

How to run an experiment

Pick a book (the equity curve to improve) and a benchmark.
Check one or more detectors — their params reveal inline; set the look-back and watch the raw score reshape.
Author a threshold regime from the score whose shape you like, and set your cuts.
Sweep — look-back, then cutoffs, then policy — watching the counterfactual NAV and the underwater drawdown; you want less drawdown without giving back too much compounding.
Compose two good signals with an operator, then re-check on other books and episodes.

Guardrail: judge a setting by improvement that holds across books and episodes at a sensible firing rate. One best-looking curve on one book is the overfitting trap.

Panels

Cumulative equity & drawdown: Book vs benchmark (log growth-of-1) over a 0-to-max underwater drawdown. Your starting point.
Detector continuous score: The signal. Channel (stress / quality / confidence / any block) and raw vs percentile. Threshold guides + state bands from the leading regime.
What it would have done to exposure: Top: the gross step. Bottom: each counterfactual NAV vs the raw hold-through baseline.
Rolling statistic: Rolling correlation between two channels (or a detector vs book drawdown), over the Corr window — a display stat, not a model dimension.

Controls

Book / Benchmark: Typeahead book (filter by name or run-date); bench = equal_weight, the book's own benchmark, or any ticker.
Detectors + ⚙: Multi-select; checking a detector auto-reveals its numeric params — the score's own dimensions (lookback, thresholds…) — and edits live-re-query (Lever 1).
Composite: Nest the checked detectors under a combination (worst_of, best_of, …).
Threshold regime: Lever 2 — author a state from a score's own cutoffs.
Policy: Lever 3 — the good/degraded/bad gross levels.
Channel / Scale / Pct window: Which signal; raw vs percentile; and the look-back the percentile rank is measured over (default 252d, tunable — not a buried constant).
Corr window / Episode: Display-only window for the rolling-correlation panel (touches no detector or state); jump to GFC / COVID / 2022 / Aug-2015.

Glossary

raw / hold-through: The book at gross=1 — the counterfactual baseline.
gross: Fraction in equities (1.0 fully invested, 0.15 = 85% defensive).
counterfactual NAV: Raw path re-sized by the detector's gross; first-order, de-risked capital earns ≈0.
state bands: good (none) / degraded (amber) / bad (red) behind the score.
look-back: How much history a score integrates — the primary dimension of most detectors (⚙), and separately the Pct window for percentile.
percentile: Rolling rank of the raw score over the Pct window (default 252d, tunable), 0–1.
block signals: A detector's internal sub-signals (e.g. turbulence vol_score).

Caveats

The detector is a universe-derived market-regime overlay, not the book's own holdings.
The counterfactual is first-order and parks de-risked capital in cash (≈0); it understates crisis protection where the real book used TLT/GLD.
Each book covers only its own run window — many start ~2010 (no GFC).
The raw path is reconstructed from the run; noisy on the rare very-low-gross day.