AI for portfolio management: from Markowitz to Reinforcement Learning | Neurons Lab (2024)

The evolution of quantitative asset management techniques with empirical evaluation and Python source code

Artificial intelligence, machine learning, big data, and other buzzwords are disrupting decision making in almost any area of finance. On theback office, machine learning is widely applied to spot anomalies in execution logs, for risk management and fraudulent transaction detection. At thefront office, AI is used for customer segmentation and support and pricing the derivatives.

But of course, the most interesting applications of AI in finance are inthe buy-sideand are related to searching the predictive signal in the noise and catching that alpha.They include but are not restricted to time series forecasting, the regime-switching detection, market segmentation, and, of course,asset portfolio management.

This article is fully devoted to the latter problem — we will review classical mathematical methods for optimizing portfolio, unsupervised, supervised machine learning approaches, reinforcement learning agents and some more exotic options. The material of this topic is tightly correlated with the inner expertise ofNeurons Labwhere I am co-founder and CTO and the course I was teaching atUCU data science summer school. As always, you can find allsource code on GitHub and the results of the experiments further in this article.

Just in case you don’t know what it is 🙂

The task title is already telling us that we’re working on the optimization problem: maximizing or minimizing some function with respect to its parameters. In our case, we would like to maximize returns while minimizing the risk with respect to the amount of money we allocate on each asset in our portfolio. In the layman terms:

We have $1M and we need to split it among different assets (or leave in cash) in such a way, that at the end of some time period this $1M grows as much as it can with minimal risk of losing the money.

Markowitz efficient frontier

If we assume, that all investors on the market are rational, avoid risks whenever possible and aim to maximize their expected returns we can plot all possible portfolios on the 2D plot with risk and expected returns as axis:

Efficient frontier visualization: https://investinganswers.com/dictionary/h/harry-markowitz

The optimal decision can be found with maximizing expected returns (calculated from the previous movements of the asset) and minimizing associated risk (as volatility of the assets). The ratio of expected return divided by risk is calledSharpe ratioand the portfolio with the maximal Sharpe ratio can be found with a ratherstandard optimization toolkit.

Custom objectives

Of course, there are a lot of different optimization criteria we might aim for. What if we even don’t care about the expected returns and just want to minimize the risk?

Or maybe we want to diversify our portfolio as most as we can?

Or probably invest only in decorrelated assets? Or maybe assure the equality of the risks in each of the assets?

Check more examples of optimization criteriahere, the most important part about it is that you candesign them yourselftaking care of your own business objectives.

Also, the portfolio optimization problem looks a lot like unsupervised learning task orrepresentation learningtask: having a set of assets we need to group them into some “clusters” based on their profitability and after allocate more funds on the most predictive ones and less on the opposite side. What algorithms could we use for this purpose?

Eigenportfolios

An illustration of different principal components of a portfolio: https://systematicedge.wordpress.com/2013/06/02/principal-component-analysis-in-portfolio-management/

One of the first unsupervised learning models you get familiar with at the machine learning class is aprincipal component analysis(PCA). It decomposes multidimensional data into the set of linearly uncorrelated variables, where the first such variable (also called principal component) is explaining the most variation in the data, and all next variables are sorted by having the maximal variance while being orthogonal to the previous variable. How we can use it for portfolio management? Many ideas are published, the mostcommon onesays that the first principal component serves as an approximation of the market, hence, choosing second and other components will giveuncorrelated to the market strategies, which is what most of the investors want.

Autoencoder risk

A visualization of the general idea behind autoencoder neural networks: https://sefiks.com/2018/03/21/autoencoder-neural-networks-for-unsupervised-learning/

PCA is a great tool, but it relies only on the linear dependence between the data axis. One of the alternatives, that allowsnon-linear dimensionality reduction, is based on neural networks — autoencoders. They can “squeeze” input data into some low-dimensional vector and after restore input from this representation. The idea of autoencoder can be exploited inmany waysfor portfolio selection, one of them is related to the evaluation of the risk carried by the particular asset: if some asset movement can’t be restored well (the predicted value differs a lot from the input in terms of, let’s say, mean squared error) from the low-dimensional representation — it’s associated with higher risk (and, possibly, higher profits). Please, see an example of this kind of portfolio in the experiments section.

Hierarchical risk parity

Geometrically, a covariance matrix of the assets in the portfolio is a complete graph (on the left), can we figure out a tree-based model that will be more optimal?

One of the optimization-based portfolio management methods is arisk parity model. It is also stated as an optimization problem, where we allocate rather the risk than the capital resources. The problem of this approach (and, actually, most of the approaches described in this article) matures when we’re working with portfolios of very huge size — if we represent connections between assets geometrically, they will be in the form of the complete graph (see image above), which is anover-complicationin the world of hierarchies. The solution lies again in unsupervised learning, but with the exploitation of thehierarchical clustering algorithms applied to the covariance matrix. After finding clusters of the assets, we can re-allocate risk over them recursively. The high-level procedure described below:

A high-level description of the hierarchical risk parity (HRP) portfolio optimization algorithm

The optimization and representation approaches look cool and legit, but they have one major drawback: they justexploit information about the pastasset movements and co-correlations without any assumptions about their future behavior. Do we agree that in the future assets will move the same as in the past? Not really! That’s why we need some ways to exploit predictions about the future as the allocation weights.

Forecasting weights

https://www.sciencedirect.com/science/article/pii/S0378437107001938

The idea is rather very straightforward: if we can use any model to predict the price movement in the future, we canuse this predictionas to the allocation weight. Of course, we need to normalize these predictions so their sum is equal to one, but this is rather a technical step and can be done or with the softmax with tuned temperature, or even with the single L1-normalization. Nice point here is, that we also can take into account associated risk of prediction if we work withBayesian machine learningor models like ARIMA. In this article, we will use the simplest forecasting algorithm: exponential smoothing, but you can easily extend it to, for example, deep learning with the use of other tutorials from my blog.

How about diving deeper and treating asset allocation not just as a one-step optimization problem, but as continuous control of the portfolio with the delayed reward? Let’s move from optimal allocation tooptimal controlterritory and in a data-driven world, it can be solved via various reinforcement learning algorithms. They don’t do predictions and don’t learn the structure of the market implicitly. They do more: directly learn the policy of changing the weights dynamically in the continuously changing market!

Deep Q-learning

An illustration of a reinforcement learning agent to decide when to enter or leave the position https://hackernoon.com/the-self-learning-quant-d3329fcc9915

The idea of Q-learning applied to portfolio management is the following: we can describe the market with some states_tand with doing some action on this market and going to the states_{t+1}we geta reward(changed value of our portfolio based on the weights we applied). The name “Q-learning” comes from theQ(s, a)function, that based on the market statesand provided actiona(in our case, it’s allocation weights) returns the expected reward. Today this function can be approximated with deep neural networks and we can train such a model to optimize for a Sharpe ratio or any other criteria, but not just for a single allocation, but for a sequence of portfolio allocations, where we will exploit trained Q function for choosing the optimal actions. If you will check the source code for details, please notice, that for continuous action space with Deep Reinforcement Learning we need touse some tricks.

Of course, asset management algorithms are not restricted to the above-mentioned ones. If we’re talking about optimization-based portfolios, it’s worth to take a look on: evolutionary algorithms that can replace classical optimization algorithms for complex target functions and huge portfolios,multicriteria optimizationthat will allow you to find the balance between a lot of different optimization targets,Litterman Portfolioand, of course, celebratedCAPM.

If you liked the idea behind unsupervised learning, I suggest taking a look atgenerative adversarial networks and variational autoencodersas the models that are showing the best results in generative modeling and factor learning today. They can help with learning market structure and hidden driving factors even more precisely.

Last but not least, optimal control direction looks like the most promising one: apart of reinforcement learning and related approaches, it’s worth to take a look at more mathematically grounded models likeStochastic Portfolio Theory.