Futures As Prelude: Bitcoin Price Forecasting From Perpetual Futures Data

Abstract

Crypto currency perpetual futures asset price changes seem to have some predictive power for short term (tens of minutes) forecasts of underlying crypto currency price changes. In this study, we apply simple vector auto-regressive models to Bitcoin spot and futures price changes and allow for a forecasting dead zone to discard forecasts near zero price change, with "near" being a tunable parameter. We have noted in-sample results that approach 70% directionally correct without extensive model refinement, with determinate forecasts occurring on a mean interval of 40 minutes or so. Such levels of accuracy and forecast intervals might be useful for guiding BTC trading. A simple out-of-sample check by a 90/10 time-ordered train/test split yielded similar accuracy as in-sample fit when the dead zone method was applied. Additional out-of-sample testing and verification of this method would probably be required before it could be used reliably in a trade recommendation system.