An open, Jev-style decision model: typed questions in, calibrated probabilities out, all read off one distribution over tomorrow's high. Every chart here is scored out of sample against the market's own price at the same timestamp. Code and raw results: physicalreasoning/isotherm. Ask it yourself: API or MCP for agents. Keyed access: join the waitlist.
Log score gain in nats per ladder. The simple pools went to zero in 2025; isotherm stays positive but decays.
Every model against the market with a 95% date-block CI. The control is trained on labels sampled from the market, so it can only win if the pipeline leaks.
Every bucket as a yes/no question: predicted probability against how often it happened.
Last 12 months. Adding NBM, the forecast behind weather.gov, no longer helps. GFS MOS still does.
isotherm minus market. Blue is better than the market, red worse; a dot marks a CI that excludes zero.
$10k bankroll, Kalshi fees. Each quarter trades the config with the best Sharpe on earlier quarters only.
Deflated Sharpe corrects for the 28 configs tried; PBO is the chance the in-sample winner lands below the median out of sample.
Believable needs DSR ≥ 0.95 and PBO ≤ 0.20. Only the daily lows 4pm cell got there, and its sealed test made $36.
Alpha against the mid versus what execution costs. Takers pay the spread; makers earn it and lose it again to adverse selection.
The configuration nested selection used most, under worse fills, higher fees, different size and participation.
Strategy and pass rule pushed before scoring; July to October 2026 scored once.
The best backtest in the project. The placebos flipped from profit to loss in the sealed weeks: the young market had corrected.
Highs profits roughly halved each month. Lows were flat.
17 Kalshi series. Weather has history, room in the spread and a free forecast. Point size: log median volume per event.
92k synthetic ladders from 55 stations Kalshi does not list. Helps when real data is scarce, slightly hurts with all of it.
Elo and starting pitchers pooled with the market, by time before first pitch.
Attention across a ladder's buckets in place of the per-bucket MLP, pre-registered. Transformer minus isotherm on the last 12 months: it needed 3 of 4 read times above zero and cleared 1, so the MLP stays.
Pooling the market with each forecast, February to August 2026.
Log score gain over the per-read MLP, last 12 months before the lockbox, 95% date-block CI; highlighted where the CI is above zero. gate: written down before it ran, needs 3 of 4 reads. explore: found by looking, so only the forward test can confirm it. control: a baseline from Merchant et al. 2026; random features with a trained head tie the MLP, and the month explains 97 to 99% of the variance in monthly gain. The weather model (§35) covers 16:00 day before only; its gate is that one read.
Market-only signals, log score gain over the mid-price market, test rows from 2026, 95% date-block CI (§39, exploration). Raising each bucket's price to a power of about 1.17 beats the mid market, but pricing buckets at their bids does most of that with no parameters: the mid counts unbid asks as probability (§41). Beyond the bid-priced market, sharpening wins only at 2pm. Sealed test against it from 2026-10-10.
Benchmark v1, 13 Kalshi series (§47). Bid vs mid: pricing each bucket at its bid instead of the mid, the §41 artefact. Outside vs in-quote: a model from the free hourly price and its recent volatility against the market at its most favourable price inside each quote; positive would be information beyond the spread. Only oil a day out clears zero; found by looking, it has a sealed test from 2026-10-12 (§48). A pre-registered test on five other daily commodity ladders passed (+0.33 a day out, §49), but the gain sits in a few stale ladders and a taker makes about nothing after fees.
What an API user receives, read off one integer distribution, last 12 months before the lockbox (§38). Noul: calibration error of P(high ≥ t) at every integer within 6°F of the forecast. Score: how much more often the 80% interval holds the high than its own probability says, and Δ CRPS against the market (positive is better).
Training on all nineteen cities instead of seven, each with its own city input: Δ log score, test 2026-07-01 to 2026-10-04, 95% date-block CI (§36). Exploration only; the pre-registered test is on January to April 2027 (§37).
A network trained on 1.16 million station-days of NBM and GFS MOS forecasts against NWS climate reports, scored at 16:00 day before on the seven cities' ladders (§35).
Gain over the market by share of training dates, pooled over reads. The MLP gains 0.003 nats from four times the data, the transformer 0.010.
Log loss by month at 16:00 day before, seven cities; lower is better. The crowd improved by about 0.15 nats in a year while EMOS-NBM stood still.
Weight on EMOS-GFS in a market+GFS pool refit on the trailing 90 days: high in summer, low in winter.
Frozen on 2026-10-06 before the first test read and scored on ladders from 2026-10-07 to 2027-04-05 at 16:00 day before; the weather model, frozen a day later, from 2026-10-09. An arm wins only if its gain over the MLP has a 95% CI above zero over the whole window.
The frozen strategy scores every city daily at 16:00 local and commits its paper trades before the outcome exists. Ledger on the shadow-ledger branch.
Rolling 30-day log score gain against the market. STOP when two consecutive windows sit fully below zero.