RegimenStatistical validation desk · does not trade · holds no funds

Methodology · the published decision rule

How Regimen decides

Every tier below is a function of the statistics, not a judgement call. The formulae, the thresholds and the controls Regimen runs against its own verdict are all here, so a result can be checked rather than believed.

Regimen answers one question about a track record: is the measured performance distinguishable from luck, and under which market conditions does it hold?

Everything below is computed from the equity curve. Nothing is taken on trust from the source's own dashboard — where the source publishes its own figures, Regimen recomputes them and reports the difference.

1. The Sharpe ratio is an estimate, not a fact

A Sharpe ratio measured over a short window carries an error bar wide enough to swallow the claim it is making. Regimen therefore never reports a Sharpe ratio alone.

Probabilistic Sharpe Ratio (PSR) — the probability that the true Sharpe exceeds a benchmark, given how long the sample is and how skewed and fat-tailed it is:

PSR(SR*) = Φ( (SR − SR*) · √(n − 1) / √(1 − γ₃·SR + ((γ₄ − 1)/4)·SR²) )

where SR is the per-period Sharpe, SR the benchmark, n the number of returns, γ₃ the sample skewness and γ₄ the sample (non-excess) kurtosis. Negative skew and fat tails reduce* the confidence, which is the correct direction and the opposite of what a bare Sharpe ratio implies.

Minimum Track Record Length (MinTRL) — how many periods the record would need before the claim could be made at the requested confidence:

MinTRL = 1 + [1 − γ₃·SR + ((γ₄ − 1)/4)·SR²] · ( Φ⁻¹(confidence) / (SR − SR*) )²

When the observed Sharpe does not exceed the benchmark, MinTRL does not exist: no amount of additional history makes that claim significant, and Regimen says so rather than returning a large number.

Deflated Sharpe Ratio (DSR) — the same statement, corrected for selection. If twenty configurations were tried and the best one is being shown, the benchmark is not zero but the expected maximum Sharpe of twenty draws:

SR*₀ = √Var(trial Sharpes) · [ (1 − γ)·Φ⁻¹(1 − 1/N) + γ·Φ⁻¹(1 − 1/(N·e)) ]

with γ the Euler–Mascheroni constant. DSR is only reported when trial Sharpes are supplied; it is never silently assumed to equal PSR.

Bootstrap interval — a stationary bootstrap (Politis & Romano, 1994) with geometric block lengths, which preserves serial dependence that an i.i.d. resample would destroy. Seeded, so a published interval reproduces exactly.

2. Evidence tiers

The tier is a published function of the statistics, not a judgement call:

TierCondition
insufficient_evidencefewer than 20 usable returns, or no dispersion at all
indistinguishable_from_luckgoverning confidence below 90%
weak90–95%, or an interval that still contains zero
supported≥95% and the bootstrap lower bound clears zero
strong≥99%, interval clears zero, and the record is at least MinTRL long

The governing confidence is the Deflated Sharpe Ratio when trial Sharpes were supplied, and the Probabilistic Sharpe Ratio otherwise.

3. Regime attribution

Each period's return is joined to the market conditions that held on the UTC date it closed on. Every factor is read with an explicit as_of date, so a bucket can only ever contain information that was knowable at the time.

Factors with a conventional scale (VIX, funding, positioning, sentiment, trend state) use fixed edges chosen in advance. Factors without one are split into terciles of their own observed range, and that is disclosed per factor, because tercile edges do depend on the sample.

The permutation test is the point. Slice a return series eight ways and the best-looking bucket will look good by chance alone. Regimen therefore compares the observed best-to-worst Sharpe spread against the spread produced by randomly reshuffling the regime labels thousands of times, and reports the resulting p-value. Without it a regime map is just a machine for finding flattering subsets.

4. Regimen attacking itself

Two controls with known answers run against every analysis:

  • Mean-centred control — the strategy's own returns with their mean subtracted. Same volatility, same shape, true Sharpe exactly zero. A correct engine must grade it near 50%. If it does not, the engine is broken and the report says so instead of hiding it.
  • Null-strategy distribution — many simulated strategies with the same length and volatility but no edge, each graded, so the real strategy's confidence can be placed as a percentile against pure luck.

5. What Regimen does not do

It does not trade, hold funds, custody keys, or sign anything. It does not audit smart contracts, score wallets or transactions, or perform security or compliance analysis. It reads a track record and reports what the evidence supports.

RegimenValidation only · no execution · no custodyBack to the desk