Forecasting of economic and financial variables is crucial for decision-making by central banks, monetary authorities, financial institutions, policy makers and international economic organisations. Many of these forecasts use 'vector autoregressive (VAR)' models, that tend to do poorly when faced with volatile changes over time, for example like those caused by Brexit. This project is aiming to use state-of-the-art computational techniques to improve these models and produce freely available software to implement these improvements.
This project received funding from the Turing-HSBC-ONS Economic Data Science Awards 2018.
Explaining the science
Vector autoregressive (VAR) models capture the relationships between multiple, evolving variables over time. The sorts of VAR models that will be developed in this project are widely used by policy makers for forecasting. For example, the Bank of England and the European Central Bank both use large VAR models as one their key forecasting models. Existing VAR models are estimated under the assumption that neither parameters nor volatilities change over time.
Consequently, when there is structural change or breaks in the sample (for example, the 2008 financial crisis or Brexit), this basic assumption may be invalid and such models are expected to perform poorly in out-of-sample forecasting, leading to potentially erroneous monetary policy decisions.
The project is part of the Turing's economic data science programme theme of 'machine learning for economic data'. The work will involve building and reviewing efficient 'vector autoregressive models' for economic and financial forecasting. These improved models will deal better with large datasets and volatile changes over time.
The performances of these models will be carefully analysed using a large UK dataset of monthly macroeconomic information, freely available from the Office for National Statistics, and financial series from the Bank of England. The research will provide a better understanding of modelling key, time-varying economic data streams such as unemployment and inflation, as well as financial series such as asset returns.
The output of the work will provide detailed guidelines of the trade-off between computational complexity and forecasting performance, and include freely available software that implements the work's findings.
Forecasting of economic and financial variables is crucial for decision-making by central banks and monetary authorities, financial institutions and institutional investors, and international economic organisations such as the IMF and the OECD.
Policy makers employ economic forecasting methods to obtain future projections of the relationships of fundamental macroeconomic variables. These forecasts are typically used to make important economic policy decisions. For example, the Monetary Policy Committee of the Bank of England uses the Bank of England model-based forecasts in order to publish the Inflation Report and decide on the short run nominal interest rate; a decision that has vast societal implications through rates on credit cards, loans and mortgage payments.
By addressing the failure of standard fixed parameter statistical models currently used by these institutions and policy makers, and generating freely available software, significant impact will be generated through improved forecasting performance.