OSD / Analysis

Segment Explorer — bootstrap 95% CIs on bet ROI by segment

What this is. Pulls all settled positions for the authenticated user and computes ROI per segment with bootstrap 95% confidence intervals. Green rows are statistically profitable; red rows are statistically leaking. Rows where the CI crosses zero are unflagged — there's not enough evidence to call them either way. This is the no-model view — exists alongside the local SHAP explorer for hypothesis-generation, but the segment CIs here don't depend on any model fit and are the more trustworthy lens for "where is ROI leaking?"

Filters

Min bets per segment

Lookback (days, 0 = all-time)

Bootstrap CI level

Awaiting auth + load.

Settled bets

—

Mean ROI

—

Profitable segments

—

Leaking segments

—

Segments scanned

—

Axis	Value	n	ROI	95% CI	Status
Click Load positions to fetch settled bets and compute segments.

Methodology & caveats

Computation. ROI per bet = amountWonOrLost / stake. Pushes count as ROI 0. For each axis (single dimension or two-dimension combination) we compute mean ROI and a bootstrap CI by resampling the bets in that segment 1,000 times with replacement and taking the empirical percentile. Segments with fewer than the configured minimum bets are excluded.

Read. Green-flagged segments (CI strictly above 0) are statistically profitable at the chosen confidence level. Red-flagged segments (CI strictly below 0) are statistically leaking. Unflagged rows have CIs that cross zero — interpret as "evidence inconclusive at this n," not as "neutral."

Survivor bias. Every bet here passed every gate. Findings condition on that — they cannot justify relaxing a gate, only tightening one further.

Multiple testing. No Bonferroni correction is applied. With ~100 segments scanned, ~5 false positives at 95% CI are expected by chance. Treat any single flagged segment as a hypothesis to verify against domain knowledge / longer history, not as a settled finding.

Scoping idea. If you want to focus the analysis (e.g. "only live bets on NHL"), trim the lookback window and re-run. A future version may add facet filters here.