Machine Learning for Finance

Targets and baselines fixed first, point-in-time features, purged cross-validation, honest labelling and feature importance, decay monitoring and the discipline to keep the simpler model when it wins.

Free · Mac (Apple silicon) and Windows · v1.0.5 · Other platforms

Then open the Marketplace in the app, search for Machine Learning for Finance, and click Install. It goes into your own files, and works offline from then on.

Before you start: this plan is general information to help you organise decisions about money. It is not financial, investment, tax or legal advice, and it does not recommend any product, investment or deal. Rules differ between countries, so check decisions that matter with a qualified, licensed adviser in yours. Investing and trading can lose money, and raising or managing other people's money is regulated in most countries.

What it installs.

1

Pillar

1

Area

50

Projects

204

Starter tasks

🧮 Quantitative Finance & Trading

Machine Learning for Finance

Applying machine learning to financial data: low signal to noise, labelling, purged cross-validation, feature importance, model decay and when simpler models win.

  1. Prediction target and horizon written down
  2. Linear and naive baselines every model must beat
  3. Point-in-time feature table with as-of joins
  4. Purged and embargoed cross-validation setup
  5. Experiment tracking for every model run
  6. Signal to noise estimate before model choice
  7. Evaluation metrics tied to trading outcomes
  8. Compute and data budget for ML research
  9. Reproducible training pipeline with fixed seeds
  10. Risk limits and stop rule for ML-driven signals
  11. Weekly review of model runs
  12. Monthly feature drift monitor
  13. Prediction decay watch for deployed models
  14. Quarterly retraining with champion and challenger
  15. Financial ML paper reading queue
  16. Label and target refresh check
  17. Hyperparameter search budget per idea
  18. Shared feature library maintenance
  19. Model inventory with owners and status
  20. Leakage incident log and post-mortems
  21. Triple-barrier and fixed-horizon labelling practice
  22. Sample weights for overlapping labels
  23. Fractional differencing for stationary features
  24. Feature importance with MDI, MDA and SHAP
  25. Gradient boosted trees on cross-sectional returns
  26. Regularised linear models and shrinkage
  27. Probability calibration and capped bet sizing
  28. Sequence models for financial time series
  29. Meta-labelling on top of a primary rule
  30. Loss function choice for ranking or regression
  31. Choosing an ML research stack
  32. Turnover and cost aware model outputs
  33. Combinatorial purged cross-validation trial
  34. Feature pruning by stability across folds
  35. Simpler model retention test
  36. Model go-live review with a written decision
  37. Research presentation of an ML signal
  38. Annual review of every model in use
  39. Workshop or conference paper on an ML result
  40. ML finance project write-up for a portfolio
  41. Adapting from tech ML to financial data
  42. Solo ML research on a laptop and free data
  43. Handoff spec from research model to production
  44. Review standards for a small ML research team
  45. Explaining a model to a portfolio manager
  46. Mentoring a junior researcher on financial ML
  47. Text and news features with language models
  48. Regime detection with unsupervised methods
  49. Reinforcement learning claims under scrutiny
  50. Online learning and concept drift handling

Where it starts.

Each project arrives with its purpose, milestones and first tasks written out. Here is how the first three read.

Prediction target and horizon written down

Most failed machine learning projects in finance never said precisely what they were predicting: next-day return, five-day sign, rank within the sector, or volatility-scaled excess return over a benchmark. Writing the target, the horizon, the universe and the moment each prediction is made, before any model is trained, stops the target drifting towards whatever the first model happens to fit.

First tasks

  • Write the target quantity and horizon in one sentence
  • State the exact timestamp at which each prediction is made
  • Name the decision the prediction will change
  • Commit the target definition beside the research code

Linear and naive baselines every model must beat

A gradient boosted model that beats nothing is not evidence of anything. Building three baselines first (a constant or historical mean, a single well-known signal, and a regularised linear model on the same features) gives every later model a bar to clear, and very often the linear model turns out to be the one worth keeping.

First tasks

  • Score a naive mean or zero prediction on the validation folds
  • Score one well-known single signal on the same folds
  • Fit a ridge or lasso model on the full feature set
  • Save the baseline table where every model report can reference it

Point-in-time feature table with as-of joins

Features built from data that was revised, restated or published late are the commonest source of leakage in financial machine learning. A feature table keyed by entity and the time the information was actually available, built with as-of joins and explicit reporting lags, means a model can only learn from what a trader could have known.

First tasks

  • List every feature source with its publication lag
  • Rebuild the feature table keyed by availability timestamp
  • Replace plain merges with as-of joins on slower sources
  • Add a test that fails on any future-stamped feature
  • Run the future-stamp test on the live feature table @recurring(weekly:mon)

Make it yours.

Install Machine Learning for Finance, then tell Aurum the date you're aiming for. It checks whether the plan fits the weeks you actually have.

Free · Mac (Apple silicon) and Windows · v1.0.5 · Other platforms

Blueprints are general information to help you organise your own work. They are not medical, financial, legal or other professional advice. See the Terms.