What I Learned Trying to Predict the Price of Cryptocurrencies
Chief Scientist, Managing Partner at Invector Labs. CTO at IntoTheBlock. Angel Investor, Writer, Boa
A few days ago, I presented a webinar about price predictions for cryptocurrencies. The webinar summarized some of the lessons we have learned building prediction models for crypto-assets in the IntoTheBlock platform. We have a lot of interesting IP and research coming out in this area but I wanted to summarize some key ideas that can result helpful if you are intrigued by the idea of predicting the price of crypto-assets.
Here are some interesting ideas:
1)Cryptocurrency price predictions is a solvable problem but not by a single approach and definitely not for all market conditions.
As the great British statistician George E. P. Box
once said, “essentially, all models are wrong, but some are useful”. This is specially true when comes to complex entities like financial markets. In the case of crypto-assets, it is definitely possible to predict price movements in cryptocurrencies but no single model is going to be effective across all market conditions. Always assume that, eventually, your models are going to fail and look for alternative.
2)There are two fundamental ways to think about prediction: asset-based or factor-based.
If you are thinking about predicting the price of Bitcoin, then you are following an asset based strategy. Alternatively, factor-based strategies focused on predicting a specific characteristics such as value or momentum across a pool of assets.
3)There are three fundamental technical approaches to tackle crypto asset predictions.
Most predictive models for capital markets, in general, and specifically crypto-assets can be grouped in the following categories: time-series forecasting, traditional machine learning and deep learning methods. Time-series forecasting methods such as ARIMA or Prophet focus on predicting a specific variable based on known time-series attributes. Machine learning methods such as linear regression or decision trees have been at the center of predictive models in capital markets for the last decade. Finally, the new school of deep learning proposes deep neural network methods for uncovering non-linear relationships between variables that can lead to price predictions.
4)Time series forecasting methods are easy to implement but not very resilient.
Throughout our experiments, we tested different time series methods such as ARIMA, DeepAR+ or Facebook’s Prophet. The results led us to believe that these type of methods haven’t been designed for complex environments such as capital markets. They are incredibly easy to implement but showed some very poor resiliency to market variations which are common in crypto. Furthermore, one of the biggest limitations of time series methods is that they rely on a small and fixed number of predictors which proven to be insufficient to describe the behavior of crypto-assets.
5)Traditional machine learning models showed poor generalization capabilities
Methods such as linear regression and decision trees have been at the front and center of quant research in capital markets. From that perspective, there is a lot of research available that can be applied to the crypto space. However, given the unusual behavior of crypto markets, we discovered that most traditional machine learning models have trouble generalizing knowledge and are very prompt to underfit.
6)Deep learning models are hard to interpret but can perform well in complex market conditions.
Deep neural networks are not exactly new but their mainstream adoption has only been possible in the last few years. In that sense, practical implementations of these models are relatively nascent. In the case of crypto markets, we discovered that deep learning models can achieve decent levels of performance when comes to predictions. However, its near impossible to interpret what these models are doing internally given its complexity and they are definitely challenging to implement.
7)There are some very interesting challenges that are not present in capital markets.
Predictive models for crypto assets encounter many challenges that are not present in traditional capital markets. From fake volumes, wash trading to the poor quality of many APIs and datasets, there is a lot of infrastructure work that needs to accompany any predictive efforts in the crypto space. Additionally many of the models included in research papers haven’t really been tested in real world markets and certainly not in crypto.
8)Plenty of challenges but also exciting opportunities.
I hope these notes provide some perspectives about the practical challenges and possibilities of predictive models for crypto assets. We will continue publishing our research and results in this space and would love to get your feedback.
Subscribe to get your daily round-up of top tech stories!