Release Notes — Daily Dew Point

April 12, 2026

Latest

Reversal Risk panel — A new section below today's prediction shows how likely the current pick is to shift before settlement, and whether conditions support an actual high that blows past all forecasts. Backed by historical data from similar setups and updated throughout the day.
Heating window mode — Between 10 AM and 3 PM local time, when temperature peaks typically occur, the dashboard automatically increases its refresh rate and signals that it's watching closely. Observations update every 2 minutes during this window instead of every 5.
Live atmospheric readings — The atmospheric signals panel (solar irradiance, cloud cover, pressure-level temperatures, wind) now shows real-time readings updated every 15 minutes alongside the daily forecast, so you can see what's actually happening rather than just what was predicted overnight.
Faster model response to warming conditions — The model previously only recomputed mid-day when observed temperatures ran cold or slow. It now also triggers a fresh prediction when conditions run warmer than expected — catching potential upside flips before they show up in market prices.

Corrupted observation guard — The NWS weather station API occasionally returns stale warm-afternoon readings mixed in with overnight data, which could cause the model to predict unrealistically high temperatures in the middle of the night. A time-aware filter now detects and discards these physically impossible swings before they affect predictions.
Improved model self-correction — The ML model tracks its own recent prediction errors to detect and correct systematic bias (e.g., consistently running warm in spring). A missing dependency had silently disabled this since launch. Now restored — the model will begin adapting to its own error patterns starting with today's retrain.

Live weather observations — Predictions now incorporate real-time weather station data throughout the day, not just forecasts. As conditions change, the model adjusts.
Intraday prediction updates — Both today's and tomorrow's predictions now refresh every 30 minutes with the latest data, instead of locking after the morning run.
Expanded model inputs — The ML model now uses additional real-time signals including observed temperature, wind, and sky conditions to improve accuracy.

Overnight temperature advisory — Dashboard now warns when overnight temperatures may exceed the next day's forecast, helping you anticipate when the daily high could be set overnight rather than during the day.
Daily model retraining — The ML model now retrains every morning with the previous day's data, instead of weekly. Predictions adapt faster to changing weather patterns.
Enhanced prediction model — Added new atmospheric features to improve accuracy on days with sharp temperature swings.
WIN/MISS tracking — ML bucket predictions are now scored against actual results daily, building a transparent track record.
Improved Kalshi integration — Live market data now loads reliably with updated API support.

Earlier this month

ML bucket predictions — Launched machine learning model that predicts the most likely 2-degree Kalshi bucket for today and tomorrow.
Top-2 bucket display — Shows the two most likely buckets with confidence percentages, giving you a realistic view of the odds.
Bet signal indicators — STRONG BET, LEAN, and SKIP labels based on model confidence.
Los Angeles support — Added LA (LAX) as a second city with full forecast tracking and ML predictions.
Server-side ensemble learning — Forecast weights now learn from historical accuracy and are shared across all users.

Foundation

Forecast tracking — Began logging every NWS and AccuWeather forecast update with timestamps.
Bias correction — Built weighted ensemble that adjusts forecasts based on historical accuracy.
Kalshi market display — Live market leaders and settlement results.
Historical accuracy stats — Bucket hit rate and forecast accuracy metrics.

Bug Fixes

Fixed false "Live data exceeds prediction" alerts appearing in the early morning hours (1–3 AM). A weather station field that spans midnight was pulling in the previous evening's warm readings and incorrectly flagging them as today's observed high.