Regimen

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

curl -s https://regimen-nu.vercel.app/api/v1/evaluate \

-H 'content-type: application/json' \ -d '{ "selector": { "source": "inline", "trackRecord": { "label": "demo", "equity": [ {"t": "2026-06-01", "equity": 10000}, {"t": "2026-06-02", "equity": 10180}, {"t": "2026-06-03", "equity": 10090}, {"t": "2026-06-04", "equity": 10310}, {"t": "2026-06-05", "equity": 10240}, {"t": "2026-06-06", "equity": 10450} ] } } }'

That curve is deliberately too short, and Regimen says so rather than producing a number. verification/README.md has a full-length example that produces a graded verdict, plus the health and deployment-proof checks.

Two surfaces, one engine

REST

Every response is { data, meta } or { error, meta }, where meta carries a request id, the build commit, and which credential mode served the request. Errors carry a stable machine-readable code, a retryable flag, and a remedy written to be actionable by an agent rather than a human reading a stack trace.

MCP

Connect any MCP client to https://regimen-nu.vercel.app/mcp over Streamable HTTP. No authentication is needed for the inline source. For a client configured by file:

{ "mcpServers": { "regimen": { "url": "https://regimen-nu.vercel.app/mcp" } } }

To inspect it interactively: npx @modelcontextprotocol/inspector and point it at the same URL.

Tools — regimen_evaluate_track_record, regimen_regime_map, regimen_self_attack, regimen_describe_factors. Each advertises an outputSchema and returns validated structuredContent; each is annotated readOnlyHint because nothing here writes, trades or signs; each takes a detail switch so an agent can ask for the verdict and its reasons rather than every bucket.

Resources — regimen://methodology (the statistics, in full), regimen://evidence-tiers (the exact grading thresholds), and the template regimen://factor/{key}, whose key argument supports completion/complete.

Prompt — validate_strategy, the full review in the right order, with instructions not to lead with the annualised Sharpe.

Architecture

                    REST  /api/v1/*            MCP  /mcp

│ │ └────────────┬───────────┘ │ engine/ significance · regime · self-attack │ stats/ PSR · DSR · MinTRL · bootstrap · conditional attribution · Wilson │ sources/ ── adapter interface ──┐ │ │ olaxbt-nexus inline (18 tools, point-in-time) (bring your own curve)

The engine is written against a domain model — a track record, a regime series — and never against a vendor's response shape. A data source is a thin adapter that produces those two things. That is why the same analysis serves an OlaXBT Nexus strategy and a curve pasted in from a spreadsheet, and why adding a venue is an adapter rather than a rewrite.

OlaXBT Nexus is the live data source. The adapter reads the strategy's equity curve, trades and published metrics, and reads eight market-condition factors per date with an explicit as_of, which is what makes the regime attribution free of lookahead. Rate limiting and caching live in the client, not at call sites: a Builder-tier key allows 80 requests a minute and a regime map wants hundreds of point-in-time reads, so calls are paced under a token bucket and every immutable past-dated read is cached.

Testing

pnpm test          # the full suite

pnpm typecheck # strict, with noUncheckedIndexedAccess pnpm eval # the graded evaluation; writes EVAL.md

The statistics core carries 370 tests. Known-answer cases for PSR, MinTRL, the Deflated Sharpe Ratio, skewness, kurtosis, Wilson intervals and drawdown were generated independently of this implementation and carry their arithmetic in a comment. A negative-control test across 20 seeds confirms a zero-mean series clears 95% confidence on 2 of 20 runs — the nominal size — while the matching positive-mean series clears it on 20 of 20.

What the evaluation found

EVAL.md is a pre-registered graded evaluation: 82 synthetic strategies with known ground truth, scored on false-positive rate, power, correct refusals on degenerate input, calibration, and regime detection. The suite and its targets were fixed before the engine was ever run against them.

It currently fails three of its six targets, and the failures are published rather than tuned away. They are worth reading, because they are the honest limits of the method:

• False positives 5.6% (2/36), target ≤5%. Both failures are the same mechanism: a negatively-skewed return distribution on a short window whose crash component simply has not arrived yet. The non-normality correction is driven by the *sample* third and fourth moments, and at 40–60 observations those carry almost no information — so the exact pattern the correction exists to catch ("sells volatility, hasn't blown up yet") is the one it walks into. The engine now says so explicitly in its own reasoning whenever a sample is too short for those moments to mean anything. Note that 2 of 36 is also within Monte Carlo noise of the nominal 5%; the mechanism is the finding, not the overshoot.
• Restraint 66.7% (10/15), target ≥90%. In five cases a genuine but unprovable edge was graded as proven, because luck pushed the observed Sharpe far above the true one. The engine is directionally right in all five and still granting more confidence than the sample size supports.
• Regime detection 25% (1/4) at p ≤ 0.05, target ≥75%. The evaluation caught a real design flaw here: the permutation test originally ran on the best-minus-worst Sharpe spread, which sees only two buckets and inflates with the number of buckets. On a planted effect it could not separate signal from the null at all — planted spread 0.493 against a null control's 0.491. It was replaced with an observation-weighted between-bucket variance, which moved the planted cases' p-values from 0.022/0.073/0.092/0.229 to 0.010/0.054/0.069/0.275 while the unplanted controls stayed quiet (0.18–0.91). That is a material improvement and still short of the target, so the target stands as missed.

Regime false positives are 0/4 and degenerate inputs are refused 8/8.

MCP conformance

npx @modelcontextprotocol/conformance@alpha server --url

Against a local build: 36 checks pass, 23 fail. Passing: protocol negotiation, ping, tools/list, resources/list, prompts/list, and DNS-rebinding protection. Every failure is a scenario that exercises the suite's own reference-server fixtures — named test tools, sampling, logging, image and audio content, resource subscriptions — none of which Regimen implements. The raw counts are printed here rather than summarised as a pass, because a conformance claim that hides its denominator is worth nothing.

Note that the DNS-rebinding scenario reports failure against any non-localhost URL by construction