UrverkOverfitting Audit

Overfitting audit certificate

LIKELY_OVERFIT

Likely overfit

Your strategy did not clear the gate. At this sample size, and after accounting for the number of variants tried, its measured edge is not reliably distinguishable from selection luck. This is the outcome for most strategies we audit, including our own fund's. Learning it now, for the price of an audit, is far cheaper than learning it with live capital.

What each gate says about your strategy

Does your edge survive selection luck and its own estimation error?Fail

Your reading: 0.3869 (gate: 0.95)

After accounting for how many variants were tried and the noise in a Sharpe measured over this many observations, your advantage is not distinguishable from luck at the 95% confidence floor we require.

How often would this strategy look best purely by an accident of where the data is split?Fail

Your reading: 0.6143 (gate: 0.50)

Across many in-sample and out-of-sample splits, the configuration you submitted often loses its lead. That is a sign the in-sample ranking is driven by the particular split rather than a durable edge.

Did the advantage hold up out of sample?Not evaluated

Not computed. This needs a baseline series to compare against out of sample.

The detection floor on your sample

At 90 observations, the smallest edge this gate could certify at all is about +6.44 of incremental Sharpe. Your observed uplift of +2.890 is below that floor, so this sample cannot separate it from noise however the other gates read.

The floor is set by how many observations you supplied, not by your strategy. More history lowers it. A short sample cannot prove a small edge no matter how clean the numbers look.

The numbers behind the verdict

Observations	90 (periods per year: 252)
Candidate Sharpe	2.890
Baseline Sharpe	0.000 (zero skill null, no baseline supplied)
Uplift over baseline	+2.890
Selection luck haircut (E max of 40 trials)	+3.391
Deflated uplift	-0.500
Incremental DSR (confidence the edge beats baseline plus luck)	0.387
Probability true Sharpe is above 0	0.952
Sharpe standard error	±1.741
PBO (probability of backtest overfitting)	0.61

These are the raw statistics the gate ran on your data. The verdict above is a mechanical function of them at pre-registered thresholds, not a judgment call.

The honest fine print

This certificate checks whether an in-sample Sharpe survives selection-luck deflation and its own estimation error, and it does NOT prove a strategy will make money going forward. A candidate that clears every gate here has only cleared the bar of 'not statistically distinguishable from noise'; a live track record, not a backtest, is the only thing that proves forward edge.
Detection floor at n_obs=90: this gate cannot reliably tell a real edge from noise below an incremental Sharpe of ~+6.44 (the smallest uplift whose incremental DSR reaches 95%). A candidate that clears the gate below is a genuine statistical finding at this sample size; more history lowers this floor.
No baseline_returns supplied: uplift/DSR are computed against a zero-skill null (Sharpe 0), not a market benchmark, and walk-forward out-of-sample uplift could not be computed.

This certificate is statistical analysis of the return series supplied by the customer. It is not investment advice and not a recommendation to buy or sell anything.

Inputs are self-attested. Deflation cannot correct for variants a customer does not disclose, and a certificate can only be as honest as its inputs.

A passing verdict is a statistical non-rejection at a 95% confidence floor, not a promise of future profit. The gate behind this certificate is the same one Ballast, a paper-traded systematic fund, holds its own strategies to, and it has killed most of them.

Urverk. Methodology and the fund's public kill-log stand behind every verdict.