Cryptocurrency researchers from Cornell University have presented a research that aims to solve the issue of bitcoin unpredictability. In the paper that was released last week, the authors discuss that though largely uncertain, returns on bitcoin are not entirely immeasurable. They have devised a methodology that predicts short-term bitcoin price fluctuations.
For their model, the researchers used order book data from leading cryptocurrency exchange, OKCoin. They believe that short-term predictions can be made about bitcoin from studying its buying and selling market trends. The reason for choosing OKCoin is because it accounts for about 39% of the global BTC trading volume.
In addition, they have also used the actual hourly volatility data of the BTC market, collected for over a year. This volatility series provides an overview of bitcoin purchases over a given time period. The order book, on the other hand, gives detailed insights into the local behavior of the market.
To make useful predictions, the presented model combines the dependency factors of both of these types of data. The order book is converted into a ‘feature series’ that includes parameters like price spread, weighted spread, ask/bid volume, slope and depth. The idea is to use this series, along with the volatility series to develop probabilistic models for volatility forecasting.
They further make use of various models to combine time series, the natural choice for volatility series, with external features. First, they detail the use of the ARIMAX model that integrates order book data as a side factor. To overcome the shortcomings of ARIMAX, a temporal mixture model is implemented which provides the required, interpretable results.
The standard learning and evaluating approach when applied to these models, compromises the non-stationary nature of data. To avoid this, the authors propose adopting a rolling incremental learning and evaluation scheme.
In this scheme, the entire time range of data is broken down to create individual time intervals. For each of these intervals, the data contained is taken as the testing set. Data contained in the intervals preceding a testing set is taken as the training and validation set. The model is evaluated individually on each interval thus providing a prediction result for each testing data set.
In addition to this model, a number of similar models have been presented by other analysts in the past as well. Cryptocurrencies like bitcoin have gained massive popularity over the past few years. These currencies are not backed by a central authority and thus lack stability, giving rise to the need for such a predictability model.
For their model, the researchers used order book data from leading cryptocurrency exchange, OKCoin. They believe that short-term predictions can be made about bitcoin from studying its buying and selling market trends. The reason for choosing OKCoin is because it accounts for about 39% of the global BTC trading volume.
In addition, they have also used the actual hourly volatility data of the BTC market, collected for over a year. This volatility series provides an overview of bitcoin purchases over a given time period. The order book, on the other hand, gives detailed insights into the local behavior of the market.
To make useful predictions, the presented model combines the dependency factors of both of these types of data. The order book is converted into a ‘feature series’ that includes parameters like price spread, weighted spread, ask/bid volume, slope and depth. The idea is to use this series, along with the volatility series to develop probabilistic models for volatility forecasting.
They further make use of various models to combine time series, the natural choice for volatility series, with external features. First, they detail the use of the ARIMAX model that integrates order book data as a side factor. To overcome the shortcomings of ARIMAX, a temporal mixture model is implemented which provides the required, interpretable results.
The standard learning and evaluating approach when applied to these models, compromises the non-stationary nature of data. To avoid this, the authors propose adopting a rolling incremental learning and evaluation scheme.
In this scheme, the entire time range of data is broken down to create individual time intervals. For each of these intervals, the data contained is taken as the testing set. Data contained in the intervals preceding a testing set is taken as the training and validation set. The model is evaluated individually on each interval thus providing a prediction result for each testing data set.
In addition to this model, a number of similar models have been presented by other analysts in the past as well. Cryptocurrencies like bitcoin have gained massive popularity over the past few years. These currencies are not backed by a central authority and thus lack stability, giving rise to the need for such a predictability model.