Workshop programme - machine learning for finance
Most courses on this topic stop at theory. This programme is built around the point where a model produces an actual recommendation - and everything that has to be true for that recommendation to be trustworthy. The work is hands-on from the first session.
After the programme ends
Financial recommendation systems are not a solved problem. The models participants build here - collaborative filtering pipelines, feature engineering workflows, bias auditing routines - remain directly applicable as the field moves. Techniques taught in 2019 still appear in production systems today. The emphasis on understanding over memorisation is deliberate: when library APIs change, the underlying logic does not.
Getting stuck is part of the process. The question is whether there is somewhere useful to turn when it happens - and what kind of help is actually available.
Each module includes a scheduled group review where participants walk through their implementations together. This is not a lecture - it is a structured comparison of different approaches to the same problem, led by a facilitator who has seen the common failure points.
Every assignment has a corresponding worked solution with detailed inline commentary. These are released after submission deadlines so participants can compare their reasoning against a reference implementation without shortcuts during the work itself.
Submitted work receives written feedback within five business days. The feedback addresses the specific choices made - not a generic rubric score. Participants can respond with follow-up questions, and those responses are answered in the same thread.
Programme modules
Collaborative filtering, content-based methods, and hybrid architectures. Participants implement each approach from scratch on a financial transaction dataset before any library abstractions are introduced.
Module detailsBuilding meaningful input features from transaction histories, risk profiles, and market data. Covers temporal aggregation, handling sparse data, and the specific pitfalls of using financial records as ML inputs.
Module detailsMeasuring recommendation quality beyond accuracy - coverage, novelty, and fairness across demographic groups. Participants audit a pre-trained model for systematic bias and apply documented correction strategies.
Module detailsServing recommendation models through an API, tracking prediction drift over time, and building monitoring dashboards. The final assignment requires participants to detect a seeded data drift event in a live environment.
Module detailsHow this differs in practice
Financial ML courses typically end at the point where a model produces a number. This programme starts from the premise that a number without context - without an audit trail, without a deployment path, without a monitoring strategy - is not a finished product. Participants spend more time on the surrounding infrastructure than on the model itself.
Dataset provenance is tracked from raw source to model input - participants document every transformation step as part of the submission.
Regulatory context is woven into assignments - not as a separate compliance module but as a constraint that shapes model design from the start.
Peer review rounds expose participants to implementation choices they would not have made - the disagreements are where the most useful learning happens.
A question worth naming directly
This is the hesitation most visitors carry before enrolling. The programme assumes comfort with Python and a working understanding of statistics - not a graduate degree in mathematics. The first module includes a diagnostic exercise that makes the actual entry bar concrete rather than descriptive.
The entry diagnostic covers matrix operations, probability distributions, and basic gradient descent intuition. Participants who struggle with any of these receive a curated list of preparation materials - specific chapters, not general reading lists - before the first module begins.
The harder parts of this programme are not the mathematics. They are the decisions: which features to trust, how to handle missing data in a regulatory context, when a model is good enough to deploy. These require judgement that develops through practice, not prerequisite knowledge.
Participants who work through the assignments in under six hours per week consistently report that the depth suffers. The programme is designed for roughly eight to ten hours of focused work weekly. That is not a sales claim - it is a structural reality of the assignment load.
The environment around the learner
The cohort model is not incidental. Participants come from data roles in banking, insurance, and fintech - the discussions in peer review sessions reflect real institutional constraints, not hypothetical scenarios. That specificity changes the quality of feedback considerably.
Tarquin Osei-Bonsu
Data Analyst, Insurance
"The peer reviews surfaced assumptions I had been making without realising. That alone was worth the time."
Przemek Walczak
ML Engineer, Fintech
"I expected the technical content to be the main value. The cohort conversations turned out to be equally useful."
