Tomorrow's forecast
Ask the model a question
Pick a price. The model returns the probability that the AAA national average ends above it. This is the same calculation that prices each Kalshi gas contract.
Shaded area is the probability above the threshold. The curve is a Student-t distribution (heavy tails), fitted by the model from recent prices, wholesale costs and volatility.
Contract by contract
Bars show the model (blue) and the market (gray). The market column is Kalshi's last traded price, captured at an unknown time, so it can include information that arrived after the model's 8 a.m. cutoff. Large gaps near the current price are therefore not a fair comparison; the scorecard below uses matched times.
Recent history
AAA national average, last 90 days with a settled value, and the next forecast.
How good is it? An honest scorecard
How it works
Ten years of EIA retail and wholesale prices train an LSTM that forecasts next week's price change and its uncertainty.
A second LSTM uses the weekly forecast, recent AAA changes, wholesale and crude moves, and a GARCH volatility estimate to forecast tomorrow's change.
The output is a Student-t distribution. Every threshold question, and every Kalshi strike, is read off its cumulative curve. Three seeds are averaged.
Built by Applied Deep Learning team 5 (Northeastern, AAI 6640): Orville Rockwell Jr. (team lead), Yi-Ting Lin, Wenjun Chen and Yixin Tian. The scorecard comes from the team's Part 4 diagnostics.