Methodology

Anytime-Valid Evidence

E-values are non-negative evidence scores for claims that are checked repeatedly as new data arrives. They let operators inspect late filings, insider cohorts, and other live research monitors without inflating false-positive rates through optional stopping.

Why E-Values Replace Repeated P-Values

A fixed-sample p-value is valid when the sample size and analysis date were fixed before looking at the data. Midas Edge research claims are often monitored continuously. Late Form 4 cohorts, insider-signal forward returns, and 10b5-1 cohorts can refresh whenever new filings arrive, so repeatedly re-running the same p-test can overstate confidence.

An e-process is built so that, under the null hypothesis, its expected value is at most 1 at every look. Ville's inequality gives the anytime-valid bound:

P_H0(max_t E_t >= 1 / alpha) <= alpha

For alpha 0.05, the decision threshold is 20. Operators can inspect the running evidence weekly, monthly, or after each cohort refresh without changing that guarantee.

Validation Status

The late-filing claim now publishes a sequential e-value artifact alongside the classical bootstrap interval, so the admin research dashboard can compare both evidence systems during Phase 1 rollout.

Source: docs/claims.json, generated 2026-06-03.

Betting Interpretation

Each new event contributes a bounded payoff. The monitor chooses a predictable bet fraction using only data observed before that event. If the signal repeatedly moves in the expected direction, capital grows. If the signal is noise or moves against the hypothesis, capital stalls or shrinks. The running capital is the e-value.

E_t = E_(t-1) * (1 + bet_fraction_t * payoff_t)

Evidence Labels

Product surfaces show plain-language labels first and raw e-values second. The formal alpha 0.05 decision still uses E >= 20.

E-valueLabelMeaning

E < 2InconclusiveKeep monitoring.

2 <= E < 10Weak evidenceSignal is directionally consistent but below decision threshold.

10 <= E < 20Moderate evidenceEvidence is building but has not crossed alpha 0.05.

20 <= E < 100Strong evidenceAnytime-valid rejection at alpha 0.05.

E >= 100Very strong evidenceLarge safety margin above the decision threshold.

Late-Filing Monitor

Compute the factor-adjusted abnormal return for each event.
Compare it with the comparable on-time filing baseline.
Convert the signed residual into a bounded payoff.
Update the running e-value and evidence label.
Publish the latest snapshot to the claims registry and admin research dashboard.

References

Vovk and Wang, E-values: Calibration, Combination and Applications.
Howard, Ramdas, McAuliffe, and Sekhon, Time-uniform confidence sequences.
Gruenwald, de Heide, and Koolen, Safe Testing.