CryptoGAT Paper: Time Series Models vs. Crypto Forecasts

A new preprint on arXiv argues that standard time series models are not the right tool for forecasting cryptocurrency prices, and that graph-based architectures that model the network of token relationships perform better.

The paper, titled "CryptoGAT: Are Time Series Models Effective for Cryptocurrency Forecasting?", tests a range of models on hourly price data for 50 cryptocurrencies. The authors compare traditional time series approaches – LSTMs, GRUs, and Transformer-based models – against a graph attention network (GAT) that treats each token as a node and uses trading volume correlations to define edges between them.

The central finding is that the graph model consistently beats the time series models on prediction error and directional accuracy. The authors report that CryptoGAT achieves a mean absolute percentage error roughly 15% lower than the best-performing LSTM across the test set. The gap widens during periods of high volatility, when token-to-token relationships shift faster than individual price trends.

The paper's argument is that cryptocurrency markets are not driven by the same kind of serial correlation that makes time series models work for equities or FX. Instead, the authors say, crypto prices are shaped by cross-token flows – a Bitcoin rally pulls capital out of smaller altcoins, a stablecoin depeg sends traders scrambling into other tokens, and exchange-specific liquidity events ripple through the network. A model that ignores those connections, they argue, is missing the main signal.