Aberdeen Asset Management Implements Machine Learning–Based Portfolio Allocation Models in the Cloud
Improve asset allocation strategies by creating model portfolios with machine learning techniques
Use MATLAB to develop classification tree, neural network, and support vector machine models, and use MATLAB Parallel Server to run the models in the cloud
- Portfolio performance goals supported
- Processing times cut from 24 hours to 3
- Results confirmed with multiple machine learning techniques
For Professional Investors Only – Not For Use by Retail Investors or Advisers
Aberdeen Asset Management (now abrdn) is one of the largest independent asset managers in the world in terms of assets under management. The company is based in 25 countries with 37 offices, over 750 investment professionals, and around 2800 staff. Assets under management were £301.39 billion as of 30 June 2016.
Aberdeen has developed a Solutions business that advises and manages on investment strategy and portfolio construction, drawing on its own experts as well as on specialist asset class teams, to provide investment outcomes tailored to specific client needs. Aberdeen Solutions bases trade decisions and multi-asset class mandates on model portfolios. Some of these models are generated with advanced machine learning algorithms developed in MATLAB® and backtested using MATLAB Parallel Server™ in the Microsoft® Azure cloud. They provide an important input into investment decision making.
“With MATLAB we can develop prototypes to test new machine learning techniques quickly,” says Emilio Llorente-Cano, senior investment strategist at Aberdeen. “Once we’ve refined the techniques and incorporated them into our asset allocation algorithms, MATLAB Parallel Server enables us to get rapid, reliable results by running the algorithms with large financial data sets on a distributed computing cluster.”
To optimize its portfolio allocation strategies, Aberdeen needed to create model portfolios in which individual asset classes such as equities, commodities, bonds, and property are overweight or underweight compared with a benchmark. These decisions are partially based on complex relational patterns linking the behavior of factors that influence markets and their impact on future asset performance. Aberdeen wanted to apply machine learning algorithms to characterize these relationships, understand their patterns, and produce trading decisions based on them.
Aberdeen analysts needed to train and backtest the machine learning algorithms using available market data. Recognizing that the more data they had, the more evidence they would have to support their results, the group wanted to use market data stretching back more than 15 years. Backtesting with this much data on a multidimensional problem was too slow for local PCs, and they needed to speed the process using a computing cluster.
Abderdeen used MATLAB, Parallel Computing Toolbox™, and MATLAB Parallel Server to implement machine learning algorithms for asset allocation and run them in the Microsoft Azure cloud.
Working in MATLAB, Llorente-Cano and his team developed a set of classification models. Each was based on a different machine learning algorithm from Statistics and Machine Learning Toolbox™ and Deep Learning Toolbox™, including neural networks, decision trees, and support vector machines (SVMs).
They trained the models using factors such as monetary policy, corporate profits, interest rates, and implied volatilities. He accessed market data using Datafeed Toolbox™.
The team backtested the trained models on more than 15 years of historical data. The tests, which are performed repeatedly as new methods are explored and new data becomes available, took up to a full day to complete.
To speed this process, James Mann, solution architect at Aberdeen, prototyped a parallel implementation on the desktop with Parallel Computing Toolbox and then used MATLAB Parallel Server to run the parallel execution on an onsite computer cluster with 80 workers.
Later, Mann redeployed the models to the same number of workers running on Microsoft Azure virtual machines (VMs). He wrote a script that allows MATLAB users to start up the VMs in the cloud, where MATLAB Parallel Server provides the machine learning algorithms access to the workers. Once finished, the users run another script to shut down the VMs.
Llorente-Cano continues to refine machine learning models for asset allocation. He is currently using MATLAB to develop trading strategies based on econophysics-inspired change-point analysis methods as well as global optimization methods in Global Optimization Toolbox.
- Portfolio performance goals supported. “We’ve based many portfolios on the asset allocation process developed with MATLAB machine learning algorithms,” says Llorente-Cano. “These algorithms help us determine whether portfolios will be overweight or underweight compared with our benchmarks.”
- Processing times cut from 24 hours to 3. “Our processing times went from 24 hours to 3 when we started running on the Azure cloud with MATLAB Distributing Computing Server,” notes Mann. “Because the job scheduler is integrated into MATLAB, it’s easy to take advantage of parallel computing just by opening a pool and using
- Results confirmed with multiple machine learning techniques. “We believe that different approaches to learning bring different types of knowledge,” says Llorente-Cano. “With MATLAB, we present the same data to neural networks, SVMs, and classification trees, and it gives us a great deal of confidence when these different models come to the same trading decision.”
For Professional Investors Only – Not For Use by Retail Investors or Advisers
The above marketing document is strictly for information purposes only and should not be considered as an offer, investment recommendation, or solicitation, to deal in any of the investments or funds mentioned herein and does not constitute investment research as defined under EU Directive 2003/125/EC. Aberdeen Asset Managers Limited (‘Aberdeen’) does not warrant the accuracy, adequacy or completeness of the information and materials contained in this document and expressly disclaims liability for errors or omissions in such information and materials.Any research or analysis used in the preparation of this document has been procured by Aberdeen for its own use and may have been acted on for its own purpose. The results thus obtained are made available only coincidentally and the information is not guaranteed as to its accuracy. Some of the information in this document may contain projections or other forward looking statements regarding future events or future financial performance of countries, markets or companies. These statements are only predictions and actual events or results may differ materially. The reader must make their own assessment of the relevance, accuracy and adequacy of the information contained in this document and make such independent investigations, as they may consider necessary or appropriate for the purpose of such assessment. Any opinion or estimate contained in this document is made on a general basis and is not to be relied on by the reader as advice. Neither Aberdeen nor any of its employees, associated group companies or agents have given any consideration to nor have they or any of them made any investigation of the investment objectives, financial situation or particular need of the reader, any specific person or group of persons. Accordingly, no warranty whatsoever is given and no liability whatsoever is accepted for any loss arising whether directly or indirectly as a result of the reader, any person or group of persons acting on any information, opinion or estimate contained in this document. Aberdeen reserves the right to make changes and corrections to any information in this document at any time, without notice.Issued by Aberdeen Asset Managers Limited. Authorised and regulated by the Financial Conduct Authority in the United Kingdom.