As part of a research project about the modeling of time series, Saffron investigated various methods of forecasting financial time series. The Bitcoin market was chosen as it is not dominated by traders with low-latency connections as the stock market is. Additionally, Bitcoin's volatility makes it an interesting market for trading.
While this project was only meant for research on time series in general, our implementation successfully predicted whether the price of Bitcoin will go up or down in the next 20 minutes with average accuracy between 60% and 80% on certain data sets from the Bitstamp exchange. This should not come as a surprise, as papers on the topic have shown the ability to triple an investment after 3 months of automated trading.
The core approach was based on the idea that the current state of the market is similar to some previous state. That is, that history repeats itself. This approach was implemented with a rolling window cross-correlation technique.
As can be seen in the these illustrations, many similar historical states could be identified for a single present state. These states are ranked by similarity to the present (lower is better).
We believe that the largest potential for future work lies in the integration of social factors together with the existing financial factors. Social events often precede shifts in the market, giving this technique considerable potential for forecasting.