Quant / 03

Automated factor research

Research methods from the literature are organised into factors and operators for automated expression search. Candidates are evaluated under fixed settings for predictive information, signal redundancy and trading costs. Their expressions, test conditions and evaluation records are retained for selection and re-evaluation.

Search seeds re-evaluated under consistent settings: 90
Methods & tools: DEAP · Strongly typed GP · HPC · Output Registry

Research visuals

Automated factor research workflow · The diagram outlines a workflow in which a research controller selects factors, operators, expression structures and parameters, generates candidates and distributes experiments across different types of compute nodes. Results are registered for independent out-of-sample testing, risk analysis and robustness evaluation, with the findings informing subsequent searches. — **Automated factor research workflow**
The diagram outlines a workflow in which a research controller selects factors, operators, expression structures and parameters, generates candidates and distributes experiments across different types of compute nodes. Results are registered for independent out-of-sample testing, risk analysis and robustness evaluation, with the findings informing subsequent searches.
View full image ↗

Genetic search convergence · Combo30 retains records from 78 search seeds, each with a population of 480. The curves show the best training fitness in each run: 67 stopped on convergence, 10 reached the generation limit, and one was stopped manually at generation 52. These are training-search and convergence diagnostics; they do not measure out-of-sample performance. — **Genetic search convergence**
Combo30 retains records from 78 search seeds, each with a population of 480. The curves show the best training fitness in each run: 67 stopped on convergence, 10 reached the generation limit, and one was stopped manually at generation 52. These are training-search and convergence diagnostics; they do not measure out-of-sample performance.
View full image ↗

Compute scheduling across nodes · The figure records Combo30 scheduling from 27 August to 7 September 2026. Lanes show staged use of Tianji, Huawei ARM, PolyU H07, SCNet, GCP, C2120, A27, DC17 and Desktop. Dashed lines connect checkpoints and task handoffs, showing how experiments moved as budgets, runtime limits, node availability and task completion changed. — **Compute scheduling across nodes**
The figure records Combo30 scheduling from 27 August to 7 September 2026. Lanes show staged use of Tianji, Huawei ARM, PolyU H07, SCNet, GCP, C2120, A27, DC17 and Desktop. Dashed lines connect checkpoints and task handoffs, showing how experiments moved as budgets, runtime limits, node availability and task completion changed.
View full image ↗