In algorithmic trading in cryptocurrency markets its always key to have an edge in the markets since the space is moving fast and adopting to the latest trends in technology.
Machine Learning algorithms can help spot patterns in historical data, model market behaviour, learn how a market behaves – both on a micro and macro level – and generate alpha for algorithmic trading companies.
Generally, asset prices follow a random walk and therefore they aren’t predictable. Especially, scientific literature has shown that past data – in whatever way it is used, modelled or interpreted – is not a good indicator to predict asset prices.
With that in mind a lot (in fact the majority) of existing frameworks, indicators or models fail in their use-case. Both the Random Walk Theory and the Efficient Market Hypothesis make clear that past events and data cannot be used to predict future events.
What both theories say is that in order to give a to a certain degree reliable future prediction one has to take current data into consideration, rather than historical data. The implication of the above is that typical models currently used in both quantitative research and algorithmic trading software is only suitable to some extent.
Why is machine-learning disrupting stochastic approaches to financial modelling?
Machine Learning and AI are rapidly evolving and are no-longer some niche scientific thing. Often used as recommendation engines, pattern recognition techniques, optimization engines or image detection system, machine learning has quite a specific field of application.
However, to utilize the capabilities of machine learning in financial market modelling and algorithmic trading in crypto it is necessary to combine it with a behavioural model since the agents in this particular use-case are not machine-driven. When it comes to machine learning one might suggest to use supervised learning as there is a set of order-book data available, however the data set that one might get from a given market to can never be optimal – there are inefficiencies in order placement, other agents that are not acting in the optimal way or other complications.
Having to produce optimal data sets for the supervised learning algorithms to use to train on, is inefficient and therefore not suitable for algorithmic trading in cryptocurrency markets. Training the supervised learning with the most efficient market available, also is not suitable neither as this would limit the model to an extent where it cannot outperform that given “almost”-optimal market. Reinforcement learning as an unsupervised learning technique however, is capable to learn without knowing the target label.
Therefore one can use reinforcement learning in order to learn how an order-book should look like so that price impact as described by is minimal whereas the second part of the problem indicates a minimisation problem.
Financial markets are complex as there are millions of agents (traders) interacting simultaneously within the environment. An upcoming trend in machine learning is MARL (Multi-Agent Reinforcement Learning) that also accounts for multiple agents to exist in the system.
Where is machine learning applied in algorithmic trading in crypto?
Financial forecasts as a rather broad term may vary from projecting balance sheets, estimating inflation numbers or usecases such as asset management, economic forecasting or other fields. Machine learning, despite the challenge of large models that – especially in terms of macro analysis – can be a helpful tool in decision making processes and allocating capital.
Algorithmic Arbitrage Trading
Arbitrage, the process of profiting from the difference between prices of the same asset on different exchanges or markets, is another field of potential use of machine learning. In such tasks ML models may be used to find optimal asset pairs – even in combination with liquidation algorithms to find the best arbitrage opportunity with the least transaction costs. (price impact/slippage)
In market making in limit order book markets, machine learning is a rather complex process to implement, however companies like Autowhale that focus on crypto market making, tackle that problem by modelling order-books and finding optimal liquidity distributions.
Are there market makers in crypto?
With the crypto economy evolving over time, more and more professional players such as market makers enter the space. Since the start of 2020 the number of market makers and crypto algorithmic trading companies are increasing and adding more professionalism to the space. The biggest players among them are GSR, Keyrock, Wintermute or Autowhale.
Bulk Asset Liquidations of large quantities
While for typical retail investors usually the execution of market orders doesn’t have a large impact on price and therefore also doesn’t involve a lot of transaction costs, institutional investors face the issue of paying high transaction costs.
In bulk asset liquidations transaction costs are high if market-depth at time T is not sufficient to cover the order quantity Q within a small range around the spread. In such cases one wants to find an optimal execution strategy that gets the best price for the selling/buying entity while keeping the risk of waiting costs minimal.
Such problems are typical liquidity problems and are often approached with price impact functions or with machine learning algorithms (optimization problems).
Of course these are just some highlights of the applications of machine learning in algorithmic trading in crypto and financial markets. What will surely be interesting is the development and integration of on-chain data into such solutions in the field of open and transparent blockchains.
How Does Crypto Algorithmic Trading Work?
While traditional markets face a wide variety of different regulations, licensing requirements and other responsibilities for market participants, crypto algorithmic trading offers a smaller barrier of entry. Similar to algorithmic trading companies on Wall Street, crypto algorithmic trading bots leverage complex algorithms to generate alpha and yield returns. Volatility and other factors are however, different and need to be considered when engaging in algorithmic trading in crypto markets.
Ultimately, the markets decide where and how machine learning will be used in financial markets, however the science and evidence so far clearly lays out that the future of algorithmic trading, quant trading desks & investments will – at least to some degree – be machine learning driven especially as science and private companies come up with more solutions to existing limitations of machine learning and better data.
None of the content above is financial advise and is for educational purposes only. Find more content on algorithmic trading software, crypto market making and market microstructure on Autowhale’s blog.