My biggest Numerai gains came after fixing one subtle preprocessing mismatch between backtests and live inference. Once training and deployment paths matched exactly, validation drift dropped and live MMC became far more stable.

Key Takeaways

  • Process parity beats optimistic offline metrics.
  • Era-wise training design matters for realistic Numerai behavior.
  • Positive MMC requires differentiated signal, not just decent raw accuracy.
Numerai tournament banner

Numerai tournament context for the 2025 season recap.

Numerai 2025

The Results

I finished Season 2025 ranked #24 overall and #4 in MMC (Meta Model Contribution), which earned the title of Numerai Master.

You can track my live performance and the current standings here:

The Breakthrough: A Preprocessing Bug

Early in the year I was stuck. My models had great cross-validation scores, but their live performance was all over the place.

The cause was a subtle bug in the data preprocessing. I only found it by comparing backtest results against live results round by round. That was the turning point: it pushed me to stop trusting good offline metrics and start trusting process parity.

My Solution: Mirror the Live Setup

This led to the approach that worked best for me: train the model as closely as possible to how it’s actually used in live deployment.

In the Numerai tournament, the model only ever sees one “era” (a snapshot of the market) at a time during live inference. Most out-of-the-box solutions fail because they let the model see context during training that won’t exist in production.

I restructured my pipeline to:

  • Treat each era as an independent unit.
  • Apply feature neutralization and ensembling at training time exactly as they occur at inference time.
  • Use loss functions specifically tailored to rank-based correlation.

By closing the gap between the “lab” and the “market,” I reduced the drift between validation and live performance, leading to much more stable returns and higher MMC.

A Word of Advice: The MMC Trap

If you want to try Numerai, do — it’s one of the more interesting ML challenges out there. Just keep in mind that most easy, out-of-the-box solutions are already widely used.

That matters because of MMC (Meta Model Contribution), the primary reward metric. MMC measures how much your model improves the Numerai meta-model.

If your solution is good but mirrors what everyone else is doing, you aren’t adding much. Worse, if it’s decent but generally less accurate than the existing meta-model and drags the ensemble down, your MMC goes negative. To win here you don’t just need a good model; you need one that knows something the rest of the field doesn’t.