This thesis reports on the applicability of drift-diffusion and diffusion-only models to estimate future price directions from Level-I limit order book data. We fit and evaluate the probability models on a data set of 78M order book updates (1 month) of the Bitcoin - U.S. Dollar market on BitMEX, a large cryptocurrency exchange. We show that diffusion-only models outperform drift-diffusion models, which is likely a result of the limited Level-I description of the order book. Both types of models offer significant improvements over baseline models that do not incorporate order book information. On selected evaluation points leading up to price changes, the best-performing model obtains an RMS direction prediction error of 9.18 - 30.52%. A model-based price direction classifier obtains an average classification precision on the same points in the range of 93.83 - 99.64%. We finally demonstrate effectiveness of the classifier in a simple trading strategy, resulting in an average return of 0.062% per trade and win rate of 69.12% over 38K trades. A realistic implementation of the strategy will require further analysis in price move sizes and/or limit order execution.

Additional Metadata
Keywords Drift-diffusion, cryptocurrency, exit problems, high-frequency trading, limit order books, market microstructure, option valuation.
Thesis Advisor Vermeulen, S.H.L.C.G.
Persistent URL
Series Econometrie
Jong, J.M. de. (2020, February 27). Order Book Dynamics - Two-Dimensional Exit Problems on a Cryptocurrency Exchange. Econometrie. Retrieved from