Educational Analytics: OULAD Click-Stream Forecasting
Forecasting daily student activity on the Open University learning platform with SARIMA, ARIMAX, Prophet and a custom 1D CNN. The CNN reached MAE 0.20 on log-scale clicks on a chronological hold-out.
400+ MB
Source data
808
Days in seasonal model
4
Models tested
0.20
CNN hold-out MAE (log)
Hypothesis
How well can classical time-series models and a small 1D CNN forecast daily student activity on the Open University learning platform?
- Problem
- Use of an online learning platform swings sharply around term starts, deadlines and holidays. Forecasting daily activity helps plan server capacity and spot unusual days.
- My role
- Team of 3 (Hardware and Software for Big Data, Module B, Federico II) with Farshad Farahtaj and Zahra Jafarinejad; I built the pipeline.
- Result
- A 1D CNN on 30-day windows reached a mean absolute error of 0.20 on log daily clicks over the last 20% of the series.
Honest limits
- Only the CNN has a real hold-out score; the other models were fitted or scored on their own training data, so they are not ranked.
- No simple baseline (for example "same as 7 days ago") was run, so 0.20 cannot be called good or bad on its own.
- Calendar dates are assumed (February and October course starts), and errors are on the log scale.
Forecasting student VLE engagement on the Open University dataset
The Open University Learning Analytics dataset (OULAD) is a public release (over 400 MB) of click-stream and assessment data across multiple UK course presentations. I aggregated it to about 2.3 years of daily totals of VLE interactions (808 days used by the seasonal model), log-transformed, and built four forecasters: SARIMA, ARIMAX, Prophet, and a custom 1D CNN.
University team project (Hardware and Software for Big Data, Federico II, July 2024) with Farshad Farahtaj and Zahra Jafarinejad; I built the pipeline.
Two different lag windows for two different model families
Statistical models (SARIMA, ARIMAX, Prophet) and the CNN want different feature shapes, so I built two parallel windows:
- 7 lag features at 1, 2, 3, 7, 14, 21, and 28 days for the seasonal ARIMA models (Prophet uses no lags), mirroring weekly and monthly seasonality.
- 30-lag input window for the 1D CNN, letting the network learn its own short-horizon shape without having to write seasonality terms by hand.
What was measured, and what was not
The 1D CNN (TensorFlow/Keras, 50 epochs, batch 32, 30-day input window) reached an MAE of 0.20 on log-scale daily clicks on a chronological hold-out.
SARIMA, ARIMAX and Prophet were fitted and plotted, but they were not scored on the same hold-out as the CNN, so the numbers are not a like-for-like comparison and no model is declared the winner. Errors are on the log scale; they are not percentage errors on raw click counts.
What each model sees and predicts
Three notebook outputs: the raw time series with anomalies flagged, Prophet's forecast against observed clicks in the original scale, and the CNN's training-data fit plus its out-of-sample predictions.



Frameworks and infrastructure
Source code on GitHub.