Economic forecasting using machine learning and alternative data
Traditionally, economic forecasting has relied upon a small data approach with a relatively small number of variables. Today we are now in a big data world. With alternative data which is continually generated to track all sorts of variables, this evolving data can be beneficial for understanding the trends in a region or the global economy.
Curating the dataset
Turnleaf Analytics curates a large dataset drawn from many sources. We combine more traditional datasets such as macroeconomic and market data with novel alternative data to enrich the dataset and improve forecast accuracy. The alternative data we use includes high frequency pollution data, which can be used as a proxy for industrial production. We also use news based data that can help us track sentiment towards inflation and other economic variables. The dataset is continually being enriched with the addition of new variables.
Enhancing economic forecasting with machine learning
Our model generates our forecasts for inflation for 33 countries, from one month to 12 months, and is updated monthly. For over two thirds of our countries, we generate a short term forecast/nowcast several days before each CPI print.
This means that we use a complex data pipeline to collect this data, clean it from missing values/outliers etc and to prepare the data for analysis. The data is then fed into a machine learning model, which handles vast amounts of data in order to find patterns within the dataset that can be leveraged for forecasting. At Turnleaf Analytics we conduct ongoing research to develop new techniques to improve our modelling.
How has our model performed?
We started publishing our forecasts in May 2022 and have performed thousands of monthly forecasts. In around 60% of the cases, our model has outperformed official benchmarks.
To illustrate the power of our forecasting, we show below , one set of forecasts for the US CPI YoY NSA, which we published in early January. Alongside our model forecast, we show the actual realised US CPI YoY NSA, which was subsequently released. We also show the expectations of the US inflation market. Our model forecast is more closely aligned to where US inflation actually ended up. For several months, our US model forecast showed more sticky inflation, well ahead of the market swinging around to the higher for longer theme.
The big news…
Turnleaf Analytics economic forecasting data is now available on Macrobond ONE.