Forecasting Cryptocurrency Prices by Xinyu Yan

Chapter 1
Introduction
1.1 Overview
In this thesis we review the various methods of forecasting cryptocurrency prices, notably bit-

coin (BTC) [Nak09]. To the best of our knowledge, previous work in this area was undertaken by
Silantyev [Sil19]. While contemporaneous and one-step-ahead prediction are of academic inter-

est, we seek to obtain tradeable results. To this end we attempt to predict the mid-price up to one
hundred ticks ahead.
We will perform the prediction on the time series data downloaded from BitMex, a crypto-currency
exchange. Before we start to train the models on the dataset, we perform exploratory data analy-

sis on the features of the trade data and market quotes from the dataset. Then we use some feature
engineering techniques to select the best derived features for the purpose of training. Some ma-

chine learning models such as linear regression, random forest and gradient boosting will be used
to predict the future tick data. Next, we print each combined out of sample prediction evaluation
line to compare the results.
For the outline of the thesis, we start introducing the concepts of the linear regression and ensem-

ble methods in chapter 2. In this chapter, random forest and gradient boosting tree are explained.
From chapter 3 to chapter 6, detailed steps are specified on how to organise the Bitcoin time-

series data to perform the prediction using the models. We carry out some statistical analysis on
the Bitcoin time series data such as bid-ask spread, bid and ask size, bid and ask price and etc in
Chapter 3. In chapter 4, we explain the methodologies on feature augmentation based on the
original features. We then extract more than 2000 features from historical order book data. In
chapter 5, we specify how we use the bespoke method to select the features from large feature
set and rank them. In that way, we try to get the most useful data for forecasting. In chapter
6, we use grid search and differential evolution to optimise the parameter space for the random
forest and gradient boosting tree forecasting respectively. Then we evaluate the performance of
the models when making decision for next 100-day prediction and make comparison. In chapter
7, we apply the feature selection by using random forest and data compression. Also, to speed
up the computational time, we use the map-reduce method to do the feature selection for large
amount of feature set. In chapter 8, we propose future directions of research. In chapter 9, we

make the conclusion of how the feature selection by using linear regression (LR) and random for-

est (RF) can affect the forecasting results by using the models like linear regression, random forest
and gradient boosting.
The whole process can be shown as figure 1.1
1.2 Forecastability
The efficient-market hypothesis (EMH) [Fam70] states that asset prices reflect all available in-

formation. A direct implication is that it is impossible to ‘beat the market’ consistently on a risk-

adjusted basis since market prices should only react to new information.
In opposition to EMH, the late Mark Joshi cited [Jos08] “the apocryphal story of the two economists
who see a ten dollar bill lying in the gutter. The first one goes to pick it up but the second one tells
him not to be silly as if it were real, someone else would have already picked it up.” He states
further that “the true moral of this story is that market efficiency only works if someone does not
believe in it—the first person who does not believe in market efficiency picks up the ten dollar
bill and thereafter it is no longer lying in the gutter.”
One should question to what degree the cryptocurrency markets are efficient. They differ from
the more traditional markets, such as equities and foreign exchange, in the high degree of frag-

mentation, stemming from the proliferation of independent spot and derivatives exchanges, such
as Binance, BitMEX, Bybit, Coinbase, Deribit, Huobi, Kraken, OKEx.

Forecasting Cryptocurrency Prices by Xinyu Yan