Lesson 5 of 12 · All lessons in this module
- Fintech and the four Vs of Big Data
- Sources of Big Data and its challenges
- Financial data science and what makes financial data hard
- AI and how machine learning learns
- Splitting data for machine learning: size, leakage and time order
- Overfitting, underfitting and model fit
- Supervised, unsupervised and deep learning
- Reinforcement learning and the four learning approaches compared
- Inside a neural network: layers, backpropagation, GANs and VAEs
- Data science: processing and visualisation
- Text analytics, NLP and data science tools
- Large language models and generative AI in finance
This lesson is part of the 2027 curriculum. You are following the 2026 curriculum, where it is not taught in this form. Switch if you are sitting the exam under the 2027 curriculum.
Lesson 5 of 12 · 12 min
Splitting data for machine learning: size, leakage and time order
The training set is the largest (roughly 60–80%), random splits keep each set representative and prevent leakage for ordinary data, but financial time series must be split in time order, with time-based splits or rolling windows that match the forecast horizon.
In short
- Rules of thumb: standard 60–80% training, 10–20% validation, 10–20% test; larger datasets 70–80 / 10–15 / 10–15; smaller datasets 60–70 / 15–20 / 15–20.
- A random split makes each subset representative and helps prevent data leakage: information from outside the training set seeping into learning.
- For time series, random splits destroy the time order and can leak future data into training. Use a time-based split or rolling-window validation instead.
- Validation and test sets should cover at least the forecast horizon (forecast six periods ahead → at least six periods each).
- A trained model can be applied to similar data, but even related data (value vs growth stocks) may need more training and fine-tuning; human judgment stays essential.
Unlock this lesson free for 7 days
Create a free account to get 7 days free: every lesson, video and flashcard, and the question bank. No card needed.