Ottawa local weather forecasting system A concise technical infographic explaining the multi-source weather data system, missing-data resilience, predictive models, held-out accuracy, human interaction and measured compute. PROJECT OVERVIEW · AUGUST 2026 OTTAWA LOCAL WEATHER A collection of target- and horizon-specific models that forecast local weather conditions and events up to 24 hours ahead. The system combines observations from many sources, tracks what is available and fresh, handles missing or broken records, and feeds reliable snapshots into predictive models. SYSTEM SCALE 2004–26 OBSERVATION ARCHIVE 22+ years of history 56 STATION NODES surface + upper atmosphere 197,892 FORECAST ISSUES point-in-time training rows 1.58M SCORED OUTCOMES across six forecast horizons FROM MANY SOURCES TO ONE FORECAST Every collection cycle records both the observations received and the sources that are late, missing or failed. SURFACE weather stations UPPER AIR profiles + soundings SOURCE LEDGER received · late · failed freshness by cadence sources continue AS-OF SNAPSHOT lags · changes · age spatial measurements missing stays missing MISSING-AWARE MODELS models accept sparse features each target routes independently a broken feed does not stop the run PUBLICATION SAFETY FIELD-LEVEL FALLBACKS carry recent values forward ≤60 min INDEPENDENT TARGETS available targets continue COMPLETE RUNS ONLY partial output cannot replace a full run FORECAST OUTPUTS · 1 · 3 · 6 · 12 · 18 · 24 HOURS Temperature · pressure · humidity · precipitation · wind · cloud · visibility · ceiling · weather events HOW THE MODELS MAKE A PREDICTION Each target and horizon selects among persistence, a regularized baseline and boosted trees on later, unseen data. PERSISTENCE assume little changes LINEAR BASELINE regularized relationship BOOSTED TREES nonlinear interactions DEPLOY strict winner only ONE REAL PATH · DEPLOYED 6-HOUR TEMPERATURE MODEL Issue: 2026-08-01 22:45 UTC · a single tree contributes one small correction HOUR COSINE > −0.129 GATINEAU CEILING 3,353 m > 665.5 m GATINEAU TEMPERATURE 3-hour lag: 26°C > 1.5°C LEAF CONTRIBUTION −0.54°C THE COMPLETE MODEL 1,728 candidate input features 1,199 trees in the median estimate 2,447 trees across median + intervals First three trees: −0.54° · −0.14° · −0.13° TOTAL RAW MODEL CORRECTION: −4.35°C HELD-OUT PERFORMANCE Temperature mean absolute error on unchanged 2026 data. Lower is better. 0.0 0.5 1.0 1.5 2.0 2.5°C 0.440.87 1.301.74 1.992.15 1h3h 6h12h 18h24h LONG-HISTORY GAIN 2.7–11.6% lower held-out MAE than the shorter-history models INTERVAL COVERAGE 94.7–95.6% observed coverage for nominal 95% intervals BUILD LEDGER HUMAN INTERACTION 72 directed messages · 3,507 words 22 50 DAY 1 · DESIGN DAY 2 · TRAIN + REFINE SELECTED MEASURED COMPUTE elapsed time; training dominates the scale LONG-HISTORY TREES 12h 21m LINEAR SUITE 19m 26s PROMOTION 14m 04s EVALUATION 3m 39s Train ≤2024 · select/calibrate 2025 · sealed test 2026 15/30-minute horizons retired where labels were insufficient.