Research
Research
A forecast is only truly useful if you can see what it rests on. That is the question behind my Master’s thesis, applied to Transformer models that forecast time series.
The problem
A time series is a sequence of measurements taken over time: energy use, temperature, traffic. Transformer models forecast this kind of data very well, but they behave like black boxes: they give a number without saying why. Yet anyone making a decision based on a forecast needs to know whether to trust it.
My approach: SEATTS
SEATTS is an explainability module attached to the output of a Transformer model. It doesn’t replace the model: it explains each forecast through four concepts, chosen to make sense to a person rather than only to a specialist.
- Level
- The baseline value the series moves around.
- Trend
- The underlying direction, up or down.
- Context
- How recent observations shape the forecast.
- Calendar
- Effects tied to the day, the week or the season.
Methods
- Models trained in PyTorch and benchmarked against the field’s reference architectures: PatchTST, Informer, Autoformer and TFT.
- SHAP, a method that measures how much each input contributes to a prediction.
- Attention maps, which show which past time steps the model relied on.
Status
- Master’s thesis
- Defended at the University of Dschang, 2026
- Paper
- In preparation
Quantitative results will be published here alongside the paper.
What I’d like to explore in a PhD
explainable AI, time-series forecasting, and models at the interface of AI and physics.