← All work

time-aware ML · uncertainty · portfolio research

Regime Residual Portfolio

A residual-alpha forecasting and portfolio-construction research pipeline built around walk-forward evaluation, validation embargoes, MC Dropout uncertainty and transaction-cost awareness.

Year
2026
Role
Research pipeline design and implementation
Walk-forward machine-learning diagram for regime residual portfolio research, with rolling training windows, uncertainty filtering and cost-aware portfolio construction.
Diagram based on the repository architecture and reported evaluation path — not a fabricated product screenshot.

Problem

Financial backtests are unusually easy to contaminate with future information or frictionless assumptions. The project focuses on whether residual-return signals can be studied with a time-aware pipeline that separates prediction from portfolio decisions.

What I built

The workflow creates residual-alpha targets and market-regime features, trains models in rolling windows with an embargo, estimates predictive uncertainty with MC Dropout, filters signals, constructs monthly portfolios and evaluates turnover and transaction costs against explicit baselines.

Research disciplineWalk-forward

Temporal leakage controls and explicit baselines are part of the evaluation design.

Engineering decisions

Where the project becomes more than a demo.

The details that matter are usually in the constraints: who is allowed to decide, what gets measured, what is kept separate, and how failure is represented.

01

Make time the first constraint.

Train and validation windows move forward and use an embargo to reduce contamination across adjacent periods.

02

Separate prediction from portfolio logic.

Model outputs are inputs to a downstream signal-filtering and allocation layer rather than being treated as trades by default.

03

Evaluate uncertainty and friction.

MC Dropout estimates predictive uncertainty, while turnover and costs keep the backtest closer to an implementable decision process.

Evidence

What the repository actually supports.

3baseline families

equal weight, momentum and ridge-based prediction

MCuncertainty

Dropout sampling used to filter uncertain signals

costsincluded

turnover and transaction friction are tracked in the evaluation

Limits

What I am not claiming.

  • The repository explicitly does not claim guaranteed alpha or future investment performance.
  • Results are sensitive to universe construction, costs, non-stationarity and data quality.
  • The main signal of the project is time-aware ML methodology rather than a single backtest number.

Signal

What this project demonstrates.

PyTorchwalk-forward evaluationtemporal leakage preventionMC Dropoutportfolio constructionresearch reproducibility

Want to inspect the implementation?

Open repository ↗Ask me about it