Skip to content

Distributed Backtesting

Under development

This optional advanced module is part of the course scaffold.

This module covers scaling backtests across many cores and machines: parallelizing parameter sweeps with Ray and Dask, aggregating results reliably, and confronting the statistical hazard that comes with the speed — a cluster that evaluates ten thousand variants overnight is a machine for manufacturing overfit strategies unless the multiple-testing accounting scales with the compute. It is for learners who have built the Part V engine and outgrown a single machine.

Topics

  • Backtesting as an embarrassingly parallel workload: sweep structure, task granularity, and shared data
  • Ray and Dask in practice: task graphs, cluster setup, and choosing between them
  • Data locality: distributing market data to workers without making I/O the bottleneck
  • Result aggregation: schema design, durable storage of sweep results, and resumable sweeps after partial failure
  • Reproducibility across workers: seeding, environment pinning, and deterministic re-runs
  • Distributed overfitting: why more compute means more multiple testing, and scaling the statistical corrections to match
  • Cost management: spot instances, checkpointing, and knowing when a sweep is not worth running