Kalshi KXMLBGAME, autonomous entry + exit: NO EDGE — and the arithmetic that forecloses it
Date: 2026-07-25 · Type: Research note — pre-registered backtest, negative result + a hard cost bound that generalises. Status: [measured — pilot, thin sample]. No trades were placed; this is research only. Artifacts: kalshi_mlb_path_extract.py · kalshi_mlb_autonomous_exit_backtest.py → results
Why this wasn't a re-chase of a refuted edge
The repo's prior Kalshi refutations are real and were respected: crypto 15M taker has no edge after fees (−$48 / 6,209 trades, now code-enforced), day-ahead weather fails its market-relative certification. But both covered other market classes and — the load-bearing point — both held to resolution, which is exactly what produces the measured "+2¢ average win / −100¢ average loss" profile. Autonomous exits change the payoff shape and had never been tested. The available tight-band data is 100% KXMLBGAME (live baseball), a class never tested, and the only data carrying the 6-second price path that makes exit testing possible at all.
The result: the gates failed
Fade-a-sharp-move entry, autonomous exit (take-profit +3¢ / stop −4¢ / 15-min timeout), fills crossing the spread both ways, Kalshi fee ceil(0.07·C·P·(1−P)) charged on both sides, market-level out-of-sample split:
| arm | n | mean net | win rate | t |
|---|---|---|---|---|
| in-sample fade | 104 | −3.82¢ | 0.23 | −6.15 |
| OOS fade | 73 | −5.37¢ | 0.16 | −6.22 |
| OOS random-entry control (same exits) | 90 | −5.59¢ | 0.00 | −34.2 |
G1 (positive net) FAIL. G2 (beats random) FAIL — the signal beats random entry by only +0.22¢, well inside noise, and both are deeply negative. G3 (power) passed. Verdict: NO EDGE.
Why — and this is the part worth keeping
The strategy was structurally guaranteed to lose before it ever ran, and the backtest merely confirmed it. Measured from the real quotes, the round-trip cost floor — median spread paid twice by crossing, plus the Kalshi fee on both sides — is:
| price band | median spread | fee ×2 | round-trip floor |
|---|---|---|---|
| 0–20¢ | 1.0¢ | 2¢ | 3.0¢ |
| 20–40¢ | 1–2¢ | 4¢ | 5–6¢ |
| 40–60¢ (where the volume is) | 1.0¢ | 4¢ | 5.0¢ |
| 60–70¢ | 3.0¢ | 4¢ | 7.0¢ |
| 80–100¢ | 1.0¢ | 2¢ | 3.0¢ |
The take-profit target was 3¢ against a 5¢ floor at mid-range. Every "winning" trade was still a net loss — which is precisely why the win rate is 0.16 while take-profits (35) and stop-losses (36) fire about equally.
Two consequences that generalise beyond this strategy:
- The fee is U-shaped, so the extremes are the CHEAPEST place to round-trip, not the most
expensive. 0.07·P·(1−P) peaks at 1.75¢/contract at P = 0.50 and vanishes at the tails, while the measured spread stays ~1¢ across bands. Floor is 3¢ at the tails vs 5¢ mid-range. (Note this does not reproduce the longshot spread premium of arXiv:2604.24366 in this venue/sample.)
- For the median market in this sample, no taker round trip can profit at all — ever. Median
total price movement across a market's whole observed life is 3¢, below the 5¢ mid-range floor. Only 27 of 79 markets move ≥10¢. Signal quality is irrelevant when the price cannot travel far enough to pay the toll.
Honest limits
Thin pilot:markets, a ~2-day window in which the two daily files cover the same markets (so the split is market-level, not temporal), and nothing resolved in-window (all active), so hold-to-resolution could not be scored as a baseline. This does not prove no edge exists in KXMLBGAME — it shows the tested strategy has none, and establishes the cost floor any future candidate must clear.
The one honestly-untested lead
Everything above is taker — crossing the spread, paying the toll twice. A maker strategy (resting limit orders) earns the spread instead of paying it, which moves the floor from about +5¢ to roughly −1¢ + fees. That inverts the sign of the dominant cost term, and it is the class the earlier no-taker-edge note explicitly left open ("a maker (limit/rebate) strategy … would need separate proof"). It is not tested here and is not claimed to work — resting orders carry adverse selection (you get filled exactly when you are wrong), which this data cannot measure without trade-direction ground truth that the Polymarket microstructure result says an order-book feed cannot reliably supply (~59% accuracy). Testing it needs fill-level data we do not have.