Everything our graded in-game signals show about how Kalshi's MLB markets behave while the game is live — win rates by inning, by entry price, by model–market gap, and the losses that come with all of it.
This is not a picks page. You will not find today's games or plays here — only the measurements.
Kalshi lists every MLB game as a KXMLBGAME event contract: a yes/no contract on the named team winning, settling at $1 if they do and $0 if they don't. While the game is live, the price trades like a market-implied win probability — a team quoted at 80¢ is being priced at roughly an 80% chance to win.
The price moves as the game does — but not at the same speed. True win probability changes the instant a run scores; the price only changes when someone actually crosses the spread with an order. In-game books are thin, nobody is obligated to keep quotes current through every pitch, and human traders reprice at human speed. The measurable delay between the game changing and the market catching up is the Kalshi lag, and we wrote a full page on it. That page is the mechanism. This page is what the mechanism produced over a season of baseball.
The measurement loop is simple. An empirical live win probability model — run differential, inning, half, and outs — reads scores every 30 seconds. Kalshi KXMLBGAME prices refresh roughly every 60 seconds alongside. When the model's probability is at least 85% and exceeds the Kalshi price of that team by at least 10 percentage points — the V2.2 filters, running since April 26, 2026 — a signal fires and both numbers are logged. At settlement, the signal grades win or loss. Across the graded set, the gap at fire runs a median of 12.5 points, a mean of 13.5, a maximum of 25 — against a median entry price of 80¢. Full model construction is on the method page.
Not most of them. All of them. The model has never once called a comeback — it is a lead-underpricing detector, not a comeback picker.
That's not a modesty setting; it's what the filters mechanically produce. A trailing team almost never models at 85%+ win probability, so the model has nothing to say about longshots. What clears the filter, over and over, is a team that is already winning while Kalshi is still quoting it below what the game state implies.
If you trade Kalshi MLB in-game — with our alerts or without them — this finding is the map. Leads are where the book is most predictably slow. When a team goes up by three in the sixth, the true probability jumps at once; the price, anchored near its pre-game level and maintained by a thin in-game crowd, credits the lead in increments. Comebacks get repriced fast and loudly — everyone sees the swing — so there is no persistent underpricing there to measure. To the extent there is an informational edge in this market at all, it is speed on existing leads: knowing the state of the game before the price reflects it. It is not vision about reversals, and nothing on this page claims otherwise.
Three cuts of the same first-signal dataset, each with its honest reading spelled out underneath. Loss rates sit next to win rates on purpose.
| Inning group | Games (n) | Win rate | Loss rate |
|---|---|---|---|
| Innings 1–3 | 40 | 80.0% | 20.0% |
| Innings 4–6 | 154 | 84.4% | 15.6% |
| Innings 7–9 | 220 | 80.5% | 19.5% |
| Extra innings | 15 | 73.3% | 26.7% |
The honest reading: extras are the weakest bucket. More than 1 in 4 extra-inning signals lost — small leads in extra innings are fragile, one swing ends the game, and n=15 is too thin to lean on anyway. Middle innings grade best; the late innings, where most signals fire, sit near the overall average. Nothing in this table says "later is safer." Both walkthroughs below came from the extras bucket.
| Entry price | Games (n) | Win rate | Loss rate |
|---|---|---|---|
| ≤ 70¢ | 31 | 67.7% | 32.3% |
| 70–80¢ | 195 | 77.4% | 22.6% |
| 80–90¢ | 203 | 87.7% | 12.3% |
The honest reading — the most important sentences on this page: the win rate largely tracks the price paid. Contracts entered at 80–90¢ won 87.7% of the time, which is close to what an 80–90¢ price already implies on its own — price is probability. That is exactly why hit rate is not profit. A high win percentage bought at high prices is not an achievement; the question that matters is whether the model's probability exceeded the price at entry, not how often the pick won. The June 14 HOU@KC loss lives in the cheapest bucket — model 89% on KC at 66¢, Houston came back 8–7. Cheaper entries lose more often, exactly as their prices say they should.
| Gap at fire | Games (n) | Win rate | Loss rate |
|---|---|---|---|
| 10–15 pp | 320 | 83.8% | 16.2% |
| 15–20 pp | 82 | 73.2% | 26.8% |
| ≥ 20 pp | 27 | 81.5% | 18.5% |
The honest reading: a wider gap is not a safer signal. The 15–20 point bucket — where the market disagrees with the model most — grades worst, with more than 1 in 4 losing. When the book refuses to move that far toward the model, part of the disagreement isn't lag at all. It's information the model cannot see: bullpen state, an injury, a lineup wrinkle. The market is slow, not blind — and any honest account of in-game trading has to hold both facts at once.
Bottom of the 10th on the South Side. The model read Detroit at 100% — the strongest output it can produce — with Kalshi quoting DET at 79¢, a 21-point gap. The White Sox then walked it off, 4–3. The signal graded a loss and went into the public log like every other loss.
This is the cleanest illustration of settlement risk we have: even a 100% model read lost. A "100%" read means the historical game states that looked like this one were essentially always converted — not that this one must be. The model computes from run differential, innings, and outs; it does not see a tiring reliever, who's due up, or a hanging slider. Nothing is decided until the contract settles.
The win — May 10, MIN @ CLE. Eleventh inning of a one-run game in Cleveland. The model read Minnesota at 100%; Kalshi was still quoting MIN at 75¢ — a 25-point gap, the widest at fire in the graded set. Nothing clever happened next. Minnesota closed out the 2–1 win the game state already implied, and the price converged at settlement. Read it as a measurement, not a prophecy: the signal didn't predict anything. It timestamped a price that hadn't reacted yet.
Notice what these two games have in common: the same kind of read, in the same kind of spot, produced opposite outcomes. The model wasn't "right" once and "wrong" once — it measured the same lag both times, and baseball did the rest. That is what an 81.6% first-signal win rate looks like from the inside: about 1 in 5.4 of these games ends like Chicago, not like Cleveland. Both games are broken down further in the May 2026 case study, and every graded signal — wins and losses — is in the sortable public log.
Team, live Kalshi price at fire, model probability, the gap in points, and the game state — score, inning, outs. The field-by-field anatomy is on the Kalshi signals page.
Alerts arrive the moment the signal fires — not delayed, not in a digest. Filter to your teams with a single command if 30 teams is too much noise.
Every signal lands in the public results log and the raw results.csv, win or lose, whether you subscribe or not.
You can audit us before paying a dollar. The free proof channel at t.me/DegenHedgeProof posts every signal locked about 10 minutes after it fires — side hidden so it can't tip anyone's hand, revealed at settlement, losses included. It has run continuously since June 13, 2026. The full product rundown, including the BTC 15-minute product that covers the same lag mechanism in crypto markets, is on the Kalshi alerts page.
The flagship product. Same signals this page is built from, delivered to your Telegram at fire, every one graded publicly — wins and losses.
Every signal can lose. 79 of the 429 graded first-signal games lost — about 1 in 5.4 — and losses cluster rather than arriving politely spaced: the longest first-signal losing streak in the dataset so far is 3 straight. At the median 80¢ entry, a single loss costs roughly what four winning contracts return, so a short cluster can erase a long stretch of wins. If a run of consecutive losses at your stake size would wreck you, the stake is too big.
The numbers on this page are measurements of one filter, over one stretch of one season, under V2.2 rules in place since April 26, 2026. Past performance never guarantees future results, and the lag itself can shrink as these markets mature. This page is education about a market mechanism for informational and entertainment purposes — it is not betting advice, and whether you act on any of it, at what size, or not at all, is entirely yours. 18+ only (some platforms require 21+). If gambling is a problem for you, call 1-800-GAMBLER.