← All courses

Live cohort · Maven

ML for Trading: From Research to Production

For people already close to the markets who want a repeatable research process, whatever your machine-learning background.

A trading strategy is not a model followed by a backtest. It is a chain of decisions about the question, market, data, timing, labels, features, validation, models, portfolio, costs, risk and production, and weak research fails at the connections between them. Start from one of nine executed case studies across ETFs, equities, futures, FX, crypto and options, or bring your own, then reconstruct and audit it, extend it, or adapt the workflow to a new problem. You leave with a cumulative research record and a defensible recommendation, which we read together one-on-one. Foundations is the grounding this assumes, and is available separately.

How it works  Twelve weeks, ten 90-minute live sessions, and a 30-minute one-on-one research review.

Enroll on Maven → See the syllabus ↓

Syllabus

Seven enrollment-time walkthroughs introduce the supplied pipeline. Twelve weeks then carry ten live sessions of ninety minutes, in weeks 1 to 10, each with an instructor-prepared second market and the week's shared experiment. Forty-six recorded units and eight assigned-reading units in nine modules carry the mechanism and are completed before the week that uses them.

  1. Week 1 Module 1

    The shared pipeline and the experiment protocol

    You are handed a working end-to-end ETF strategy, small enough to read in one sitting, and it is what every experiment in the course modifies. Week 1 sets up the study folder and findings ledger, fixes walk-forward with purge and embargo as the protocol every experiment shares and what counts as one trial, and orients you to the nine case studies and to using coding agents in the course.

    • Live session, 90 minutes
    • Second market: CME futuresan ETF is a fund with a NAV; a futures contract is a dated obligation with a roll and embedded leverage
    • This week's experimentsupplied, and everyone runs the same one
  2. Week 2 Module 2

    Market, instrument, and data construction

    What is traded, and what object could have been observed at the decision time once the clock is not the daily close. Asset class, instrument and vehicle, and the target capital that constrains them; time, tick, volume, dollar and imbalance bars; reference price, roll and back-adjustment, with point-in-time correctness read under back-adjustment.

    • Live session, 90 minutes
    • Second market: Crypto perpetuals + CME futuresbar construction, reference price, and roll are degenerate on a daily close
    • This week's experimentsupplied, and everyone runs the same one
  3. Week 3 Module 3

    Model-based features

    Foundations builds features that are computed. These are estimated: they refit, they can be wrong, and they carry their own validation burden. Fractional differencing at a chosen order, residualization and the different question it asks, GARCH and HAR and rough volatility, regime as an observable rule or a hidden state.

    • Live session, 90 minutes
    • Second market: FX pairsroughly five effective bets, with the dollar on one side of most pairs, so residualization is required
    • This week's experimentsupplied, and everyone runs the same one
  4. Week 4 Module 3

    Feature validation, labels, and mechanism testing

    Falsifying a feature's mechanism rather than confirming its correlation, with SHAP as evidence; path-dependent labels compared across markets - which one forces triple barrier, trend scanning or meta-labeling, and why - against the fixed-horizon baseline; and when a model-based feature earns its estimation cost.

    • Live session, 90 minutes
    • Second market: S&P 500 option analyticsskew, term structure, and the variance risk premium are estimated features used to trade a different instrument
    • This week's experimentthree alternatives are supplied; you pick one
  5. Week 5 Module 4

    Model capacity: deep learning and gradient boosting

    Whether ordering carries information your engineered windows do not, whether sequence architectures beat the linear baseline that often wins, whether tabular deep learning earns its runtime against gradient boosting on identical folds, and how configurations combine across families. From here you choose your own experiment.

    • Live session, 90 minutes
    • Second market: ETFs (case study)the six-family comparison in which the highest-IC family is not the highest-Sharpe family
    • This week's experimentthree alternatives are supplied; you pick one
  6. Week 6 Module 5 + 6

    Latent structure, and when the question is causal

    Recovering low-dimensional structure from a wide cross-section, then separating prediction from an identified causal effect. Factor count and purpose; the estimand and identification design; double machine learning and structural time series; and the placebo, sensitivity, and stability tests that may refute the result.

    • Live session, 90 minutes
    • Second market: US firm characteristicsa wide characteristic panel supports latent structure, while its accounting lags and survivorship make the identification assumptions concrete
    • This week's experimentthree alternatives are supplied; you pick one
  7. Week 7 Module 7

    From signal to executable order

    How predictions become positions, and how much of the result that layer owns. Which guardrails bind first; the covariance the allocator consumes and the window behind it; allocators beyond equal weight; volatility targeting; and execution algorithms as assigned reading.

    • Live session, 90 minutes
    • Second market: NASDAQ-100 microstructurethe cost floor that makes a raw signal lose money, on the most cost-favorable panel in the book
    • This week's experimentthree alternatives are supplied; you pick one
  8. Week 8 Module 7

    Risk, stress testing, and failure analysis

    Factor-model attribution of exposure, regime-triggered caps and turnover limits, value at risk against conditional value at risk and the estimator behind each, stress and reverse-stress testing, the calibrated kill switch, and regime slicing under a stated snooping discipline.

    • Live session, 90 minutes
    • Second market: S&P 500 optionsa delta-hedged panel where most predictors are significant gross and few survive costs
    • This week's experimentyou propose it
  9. Week 9 Module 8

    Experiment tracking and multiple-testing corrections

    The run log at registry scale; corrections for repeated experimentation from Bonferroni through false-discovery-rate control to learning-theory bounds; the deflated Sharpe ratio and reality check beside the Rademacher antiserum; and conformal prediction. Then one holdout candidate and its access condition are recorded.

    • Live session, 90 minutes
    • Second market: US equities panel + FX pairsa wide panel with paired-bootstrap confidence intervals applied to a weak signal
    • This week's experimentyou propose it
  10. Week 10 Module 9

    Going live

    If the declared condition is met, the holdout is opened once and read against the trial count. Then: broker and platform selection, the order lifecycle and rejection recovery, research-to-live parity tests, and a phased rollout with explicit revert conditions. The final live session tests the deployment design against crypto perpetuals, where direct venue access, continuous trading, and recoverable order state make each choice observable.

    • Live session, 90 minutes
    • Second market: Crypto perpetualsa direct, continuously open exchange makes venue choice, state recovery, research-to-live parity, and phased rollout explicit
    • This week's experimentyou design it and defend the design
  11. Week 11 Module 9

    Monitoring and drift

    Specify the rolling metric set and alert thresholds, test for data, feature, and concept drift, decide when evidence requires retraining or pausing, and read the circuit-breaker design. The capstone is drafted from the accumulated findings ledger.

    • No live sessionmonitoring and drift while the capstone is drafted
  12. Week 12 Capstone

    What you established

    Your report: which decisions actually moved the objective, ranked; what survived scrutiny and what did not; the week-10 holdout read against your trial count; the deployment plan for the specification you would put live; and the most useful next question. Assessed on the quality of the research, with reported returns read as evidence within it rather than as the target.

    • No live sessioncapstone submitted; written cohort review returned
    • All nine case studies as a set
Enroll on Maven →