← 1st Round

Foundational Models

Work comparatively well with lack of data, outperformed by DL (e.g. LSTM) otherwise.

  • TimeGPT
  • TCN
  • 1-d CNNs

Hybrid (LSTM + GRU)

Combines architectures for long/short term forecasting.

Reference: Attention-Based Transformer + GRU Model for Predicting Cryptocurrency Prices

LSTM

Reference: A Cryptocurrency Price Prediction Model using Deep Learning

Random Forest

Reference: Integrated Framework for Cryptocurrency Price Forecasting

AR + L1 + L2 Regularization

Found going through papers on energy balancing market forecasting. The market data is highly complex and easy to overfit so could be good for our use case.

Custom Anti-Overfit Model

Custom model written for DRW Crypto (performed best then). Caution: Competition was with CS data.