LEMON System Outperforms Rivals in Cryptocurrency Price Forecasting

Cryptocurrency prices are notoriously volatile, making it challenging to predict price movements. Researchers in Italy and Slovenia have developed a machine learning system called LEMON that addresses this issue by grouping related cryptocurrencies together and assigning higher weights to instances showing large price variations.

The team's method first identifies which assets behave alike using temporal clustering with the Dynamic Time Warping (DTW) distance measure. It then trains a shared model for each group, allowing the collective behavior of the cluster to inform every individual forecast within it.

LEMON uses a prequential predict-update cycle, continuously adapting to shifting market conditions by updating its model after each new instance arrives. This approach is crucial in an environment where trading patterns and trends can change rapidly.

The system's dynamic instance weighting scheme assigns higher weights to extreme values, which often mark the onset of abrupt trend changes that traders need to capture. The method uses Tukey's boxplot rule and the medcouple statistic to identify outliers and assign weights accordingly.