Findings

Write the hypothesis down,then let the data reject it.

Before an analysis I write down what I expect. These are the ones the data did not let through.

2026.10

AI weatherWith data

Pangu-Weather and GraphCast did not beat the grid operator's wind forecast

The gap has two parts: without 100 m wind, even perfect weather trails the official forecast by 0.7; forecast error adds 1.1.

Day-ahead in 2022, the same weather-to-power model driven by ECMWF HRES, Pangu-Weather and GraphCast gave MAE of 7.80%, 8.10% and 8.19% of capacity, against 5.97% for ERCOT's official forecast; every 95% interval of the difference excludes zero. After removing each source's annual mean bias, neither AI model differs significantly from HRES.

I had assumed: ~~Operational AI weather forecasts, plugged straight in, would match or beat the official wind-power forecast.~~

Evidence

Operational comparison

2026.09

Wind powerWith data

The official day-ahead forecast already matches a perfect-forecast ceiling

The remaining room is between available power and actual generation: curtailment.

ERCOT's 2024 day-ahead wind forecast has an MAE of 6.24% of capacity; my perfect-forecast ceiling built on reanalysis winds is 6.38%. The official forecast's high bias against actual generation comes mostly from curtailment.

I had assumed: ~~A model of my own could beat ERCOT's official wind-power forecast.~~

Evidence

Bias is curtailment Error by lead time

2026.09

Wind power

Seasonal coverage of the training data matters more than the model

Over a full 2024 test year, XGBoost, LSTM and a Transformer differ by less than 0.2 percentage points. Trained without autumn and winter, every model is biased low by 3.2–5.9 points (+0.5 to +1.3 when trained on a full year).

I had assumed: ~~A more complex model would bring wind-power error down.~~

Evidence

Model comparison

2026.07

Nowcasting

Which nowcasting method wins has to be tested case by case

In a stratiform case, the radar–NWP blend lost to plain extrapolation at +1 h; a storm-structure rule derived from the first two cases failed on the third.

I had assumed: ~~Blending in NWP would always beat extrapolation alone.~~

Evidence

Method comparison

2026.09

Hazard remote sensingWithin 3 days of the event

Cross-season images cannot simply be differenced

Remove the scene-wide seasonal drift and mask the valley floor before estimating damage.

For a high-mountain ice–rock avalanche, differencing images from different seasons overestimated the affected area about 20-fold.

I had assumed: ~~Subtracting before and after images outlines the damaged area.~~

Evidence

—

2026

Dust climate

Check for sensor breaks before reading a long-term trend

All 11 sub-regions step together in 2004/05, exactly at the TOMS→OMI sensor transition. Within the OMI era (2005–2024) no sub-region declines significantly.

I had assumed: ~~The East Asian aerosol index declined steadily over 2000–2024.~~

Evidence

—

2026.09

Kaggle

When local evidence is strong and the public board says otherwise by a hair, keep a final slot for it

The public leaderboard covered under 30% of the test data. The two changes I dropped because the public board rejected them turned out to be the two largest gains on the private board.

I had assumed: ~~Public leaderboard scores tell whether a change is worth keeping.~~

Evidence

—