Skip to main content
← Retour au blog

Maelstrom Negative Finding: Why Strategy Embeddings Did Not Become a Live Trading Edge

📅 2026-05-28
✍️ Strategy Arena
maelstrom negative finding machine learning brier score contextual bandit strategy embeddings

The result

Maelstrom was built to test a simple but attractive hypothesis: if we learn an embedding for each trading strategy, can downstream ML models use that compact identity vector to improve out-of-sample decisions?

The answer, after the final target ablation, is no. Not enough to promote it.

Target Baseline Brier With Maelstrom embeddings Relative change Verdict
Leviathan NN 0.233428 0.237749 -1.85% Red
Random Forest 0.250272 0.250925 -0.26% Red
Hydra-like LSTM 0.252075 0.257758 -2.25% Red
Inverse-Brier ensemble 0.247899 0.247857 +0.02% Red

The promotion criterion was strict: Brier improvement of at least 1.5% and better performance than a random embedding control. The ensemble moved by only +0.02% and lost against both the zero and random controls. That is not a usable edge.

Why this matters

Most trading AI projects only publish winners. Strategy Arena keeps the negative results because they are part of the research map. Maelstrom did learn something about strategy identity, but the learned vectors did not transfer into robust predictive value for the models that matter.

That tells us something useful: for this market/timeframe, compact strategy identity alone is not enough. The stronger path remains Monte Carlo-validated strategy cells, live drift tracking, regime filters and risk-aware allocation.

What was archived

The Maelstrom family is now a scientific archive, not a live candidate:

  • Maelstrom Bandit
  • MaelstromGated Bandit
  • Maelstrom Minimal

The artifacts, visualizations, post-mortem files and final target reports are preserved under:

/var/www/strategyarena/_archive/maelstrom_negative_finding_20260528/

No live Strategy Arena trading path was promoted from this experiment.

The useful lesson

Negative ML findings are not failures when they close a loop. Maelstrom closed one:

Strategy embeddings are interesting as research artifacts, but they did not beat simple controls across downstream targets.

So we archive the branch and keep the evidence.

⚠️ Avertissement — Cet article est publié à titre informatif et éducatif uniquement. Il ne constitue en aucun cas un conseil en investissement ou une recommandation d'achat/vente. Les performances passées ne préjugent pas des performances futures. Strategy Arena est un simulateur éducatif avec capital virtuel. Faites vos propres recherches avant toute décision d'investissement.

Cet article vous a plu ? Partagez-le

𝕏 Partager sur X ✈️ Telegram