ML research roundup for quants
A new arXiv comparison ran 918 experiments across deep-learning architectures for multi-horizon financial forecasting, while threads argue markets are a natural lab for multi-agent reinforcement learning and agent-based models. Other posts revive Santa Fe-style simulations showing how simple heterogeneous agents generate fat tails and momentum — useful methodological reference points for trading-strategy design. (x.com) (x.com) (x.com)
ModernTCN takes the top spot on the paper’s global leaderboard with a mean rank of 1.333 and a 75% first-place rate across the 24 evaluation points, while PatchTST is the nearest challenger with a mean rank of 2.000. (arxiv.org) The authors report that architecture choice explains 99.90% of raw RMSE variance versus 0.01% attributable to seed randomness, and that rankings hold across horizons even as pointwise errors amplify by roughly 2–2.5×; directional accuracy for MSE-trained models is statistically indistinguishable from 50% across the 54 model–category–horizon cells they test. (arxiv.org) Their protocol yields four explicit operational takeaways: large-kernel temporal convolutions and patch-based Transformer designs outperform others, model complexity and parameter count do not map monotonically to performance, three-seed replication is sufficient given negligible seed variance, and achieving directional forecasting will require loss-function redesign rather than off-the-shelf MSE objectives. (arxiv.org) The paper’s author published the full codebase, data splits, trained weights and notebooks to enable independent replication and hyperparameter audits on GitHub under the repository used for the experiments. (github.com) Recent ML/finance work frames markets as an explicit testbed for multi-agent RL research — for example, papers on multi-agent market-making and multi-agent limit-order-book simulations use hierarchical and competitive MARL setups to study emergent collusion, market impact, and generalization across instruments. (arxiv.org) (hal.science) Contemporary revivals of Santa Fe–style agent-based markets put runnable Python reproductions back on public repos and documentation pages, demonstrating that simple heterogeneous agent rules can reproduce fat-tailed return distributions, volatility clustering and other Cont-style stylized facts used to justify momentum and tail-risk strategies. (github.com) (santafe.edu) Several recent educational and review resources revisit the SFI-ASM’s role as a methodological reference, showing how tuning agent mutation/selection parameters and information sets produces persistent momentum and heavy tails, and providing code-level parameter files that quants can adapt for strategy microstructure experiments or stress-test scenarios. (complexityexplorer.org) (link.springer.com)
