The Helios Corn Climate Challenge doesn't score a forecast. It scores how well your climate-risk features correlate with how corn futures behave. So the features are what you submit.
If the metric is correlation, then the first thing to build is something that can measure correlation the same way the leaderboard does.
Step zero: my own copy of the metric
Before any feature work I wrote a mock of the competition's correlation metric. Each submission otherwise becomes a slow, limited experiment, and I'd end up tuning to leaderboard noise. With the mock I can loop as often as I like:
⚠️ Warning: a mock metric is only useful if it agrees with the real one. Self note: check the mock against at least one real submission score before I trust it.
Then the features
- F-score feature selection: rank every candidate climate signal by how much it explains on its own, and drop the weak ones early.
- Correlation-based weighting: combine the survivors with weights based on how strongly each one correlates with the target, so one noisy signal can't dominate.
F-score selection keeps the list short, the weighting does the combining, and the mock metric tells me if the combination is any good.
Where it's at
- ✅ Offline mock of the correlation metric
- 🔄 F-score selection
- 🔄 Signal weighting