---
name: xlow backtest TEST-segment decomposition — the +48u is one hot mid-May week
description: The backtest's strong last-40% segment is 98% May bets; its top-3 days (May 13/16/17) contribute 74% of the +48u. Driver is the Poly-cheap leg's WR jumping 26%->43% plus sub-15c longshots flipping. True edge plausibly nearer train's ~+10% than the blended +24.7%. Median week is ~flat; the mean is made by rare fat weeks.
type: project
---

Decomposed 2026-07-19 via `E:\soccer\research\xlow_test_segment_decomp.py` (rerouted-to-
Kalshi stream, same 286 bets as [[project_kalshi_only_execution]]).

**Segments are calendar-lopsided:** TRAIN = Apr 22-May 12 (n=171, +18.2u, WR gap +3.0pp).
TEST = May 12-Jun 14 but 113/115 bets are MAY (Poly feed dark from ~Jun 18). So "test"
is effectively mid/late May.

**Findings:**
1. **Day concentration:** TEST top-3 days (2026-05-13/16/17) = +35.3u of +48.0u (74%).
   One weekend carries the segment AND the "+43.7% OOS tail" headline.
2. **Not price/league mix:** avg entry px identical (0.237 vs 0.239); league mix similar
   (MLS strong in both). The change is pure win rate: gap over implied +3.0pp -> +10.2pp.
3. **Poly-cheap leg is the swing:** Kalshi-native ~equal both halves (+20.1u/+23.6u);
   Poly-cheap bets went -1.9u (WR 26%) -> +24.4u (WR 43%). Matches the "TRAIN -4.4% /
   TEST +47.4% front-loaded" noise flag already in the Kalshi-only study.
4. **Sub-15c longshots flipped:** TRAIN 1W/25 (-16.7u, the f10-floor motivation) ->
   TEST 4W/19 (+12.0u). Longshot variance both directions.

**Permutation test (2026-07-19, same session):** day-blocked shuffle (10k perms,
same-day bets kept together): P(random last-40% beats first-60% by >= the observed
+31.1pp) = **0.073**; P(poly-leg WR jump >= observed +16.4pp) = **0.060**. Both gray —
the split difference is most consistent with LUCK OF THE CALENDAR, not a real time
trend. Consequences: (a) the "+43.7% OOS tail" is not special evidence — use the
POOLED backtest (+66u/286, z+2.4) as the estimate; (b) symmetrically, there is no
evidence of a special "May regime," so forward's flat stretch needs no regime
explanation either. Fat weeks are real but randomly placed.

**How to apply:**
- Discount the blended +24.7%/+17.8% magnitude: effective evidence is ~6 calendar
  weeks, day-clustered. If real, the edge is plausibly nearer TRAIN's ~+10%/bet.
- The weekly P&L distribution is fat-tailed: median week ~flat, mean made by rare hot
  weeks. A flat forward month is NOT disconfirming by itself — but it also means
  "confirmation" at n>=100 should weight the WR-vs-implied gap (a per-bet statistic)
  over ROI (dominated by whether a fat week happened to land inside the window).
- Do NOT use this to re-tune rules (e.g. dropping the Poly-cheap leg mid-test = the
  same forking-paths trap). The [[project_xlow_rejection_criteria]] stand unchanged.