Forecasting hackathon

252 days ahead for six indices

2026MIAX master's, Instituto BME

Figure 1. The task as a diagram: each prediction feeds the next one

Summary

A team hackathon: forecast six indices day by day for 252 steps, each prediction feeding the next, scored by average RMSE. We set flat, drift and random-walk baselines first and spent the network effort on the indices that needed it. Team of 3.

The format

A one-day team hackathon: forecast six indices day by day for 252 steps, where each prediction feeds the next, scored by average RMSE.

How we played it

First a short exploration of each index's quirks, such as whether one copied another with a lag. Then flat, drift and random-walk baselines for all six, to secure the pass. Only after that did the network effort go where it was worth it, with LSTMs planned for the hardest indices and a flat line for the steady one. Each index had one owner and one results file, and a final notebook only read them to build the submission.

Figures

**Figure 2.** Correlation of index D with the other indices by lag, from 0 to 30 days

**Figure 3.** Error of the flat, drift and random-walk baselines on each index

**Figure 4.** Raw prices of the six competition indices

Built with

PythonKeras + TensorFlowLSTMpandasNumPy