Regimen MCP server

Regimen

A Sharpe ratio is an estimate. Regimen tells you whether it is a fact.

Regimen takes a trading strategy's equity curve and answers two questions an aggregate performance number cannot: is the measured edge distinguishable from luck, and in which market conditions does it actually hold?

It is built for agents as much as for people. Everything is available over a REST API and over MCP, and the most common answer it gives is that the evidence is too thin to support the claim being made. That is the product, not a failure mode.

• Live API — https://regimen-nu.vercel.app
• MCP endpoint — https://regimen-nu.vercel.app/mcp (revision 2026-07-28)
• OpenAPI — https://regimen-nu.vercel.app/api/v1/openapi.json
• MCP Registry — published as io.github.RaYYeR220/regimen (listing)
• Verify it yourself — verification/README.md · Claims ledger — CLAIMS.md · Real vs simulated — MOCKS.md · Scorecard — EVAL.md

The problem

A strategy publishes a Sharpe ratio of 2.4 over six weeks and a 58% win rate over 40 trades. Both numbers are real. Neither is evidence.

A Sharpe ratio computed from a short, skewed, fat-tailed sample carries an error bar wide enough to swallow the claim. Forty trades cannot distinguish a 58% edge from a coin. And once a strategy has been re-tuned twenty times, the best configuration looks good for the same reason the tallest of twenty random people is tall.

This is not a niche statistical objection — it is the single most common way capital is lost to a backtest. The mathematics for handling it has existed since 2012 and is almost never applied, because it requires more than dividing a mean by a standard deviation.

Here is real output from the live service, on a curve with an annualised Sharpe of 3.72:

verdict      weak
confidence   92.3% that the true Sharpe exceeds 0
95% interval [-0.033, +0.462]        zero is still inside
track record 60 periods; 78 needed for significance at 95%

A dashboard would have printed 3.72 and stopped.

What it does

1. Significance. The Probabilistic Sharpe Ratio — the probability the true Sharpe exceeds a benchmark given the sample's length, skewness and kurtosis. The Minimum Track Record Length — how many periods would be needed before the claim could be made at all. The Deflated Sharpe Ratio — the same statement corrected for how many configurations were tried first. A stationary-bootstrap confidence interval that preserves serial dependence.

2. Regime attribution. Each period's return is joined to the market conditions that held on that UTC date — volatility, funding, open interest, positioning, sentiment, trend state — read point-in-time, so nothing in a bucket could only have been known afterwards. Every factor carries a permutation test: the observed dispersion of performance across buckets is compared against the dispersion produced by randomly reshuffling the regime labels, because slicing a return series eight ways guarantees a flattering subset. Without that p-value a regime map is a data-mining machine.

3. Self-attack. Every analysis can be run against controls whose answer is known in advance: the strategy's own returns with the mean removed (true Sharpe exactly zero, so a correct engine must grade it near 50%), and a simulated population of edgeless strategies matched for length and volatility, so the real result can be placed as a percentile against pure luck. The result is published, including when a control fails.

4. Divergence check. Where a source publishes its own figures, Regimen recomputes them from the equity curve and reports the difference. Differing conventions explain most gaps, but a user quoting a dashboard deserves to know when the curve underneath says otherwise.

Try it, with no credentials

Shortened. The full README is on GitHub.