Asger covers macroeconomics forecasting of Chinese time series using a large number of prediction variables. He investigates what is the extent of improvement of forecasts when news sentiment indexes built using RavenPack data are included among the predictors. The results suggest that forecasts obtained with this method outperform univariate autoregressions and in shorter prediction horizon news indexes improve the forecasts.
How we can do better?
- We live in a time of Big Data and increased computational power.
- Practitioners and governments are interested in macroeconomics forecasting the state of economy to make.
- Investment and policy decisions.
- However there are problems with data releases – low frequency with delay.
- Typical frequency in macro dataset is monthly and quarterly, with additional annual variables.
- Ravenpack’s news data offers additional insight into the state of economy.
Problem Formulation: Macroeconomics forecasting
In terms of what we have in regards to big data in the world that tells you something about economic activity. So if you want to predict microeconomic activity, then you need all the data to be infinity precise. What we do is aggregate everything into a few macroeconomics indicators, such as interest rates, balance of payments, production and so forth.
Contribution and approach
The approach that we are going to take is anchored in the ecometrics. We work with machine learning techniques but I wanted to anchor it to the literature it came from and in that literature, what you have been doing is you take all the macroeconomic indicators that you can get access to, such as federal reserve, IMF, OCD. In our case we have around 130, then you extract the components from that, and then you tuck them into a time series model. Of course you can always use your favourite machine learning technique to do the forecasting.
This is extremely low frequency and these are very well established organisations, that have a rigid way of doing their macroeconomic data.
We are taking all the data from the macroeconomic and trying to extract some information from that. We then augment that with the data sentiments.
The datasets consist of: balance of payments, stock prices, domestic product, foreign relations, government, natural disaster, product services, foreign exchange, consumption, acquisition and mergers, assets, earnings and many more.
Looking at the common factors of datasets
Macroeconomics forecasting improvements for GDP in a three month horizon
After estimating the model, you can see the performance on the left side. Its organised in a way that positive numbers go to the right. Outperforming the benchmark and one without any macro or news and the negative ones are not outperforming. Its split into three parts. The first chunk of results is only using the macro data. The second is combining the macro variable with the lag value of the GDP. The third is only the sentiment.
Macroeconomics forecasting improvements for GDP in a six month horizon
What you see here is that the value of the news is not as big as before. Maybe this is due to the news being old.
Comparison of GDP models at three and six month horizon
If you do a combination of sentiment and the target then you get more mixed results but it is in general the case that the news is more informative than the macro variables.
Conclusion: Macroeconomics forecasting
When we do research into macroeconomics forecasting, we should take new data sources into account, we should not trust so much on the company data that you get, you have many more sources of information.
- Showed that including RavenPack Event Sentiment Scores (ESS) indexes improves the forecasts of macroeconomic time series for China.
- Approached the “Big N small T“ problem with Principal Component Analysis.
- In the 3-month prediction horizon we find that sentiment improves forecasts for Gross Domestic Product, Balance of Payments and Unemployment, but not for Exchange Rates.
- Extended the same analysis to 6- and 12-month analysis of the same variables, but the added value of ESS dataset diminishes with the length of the prediction horizon.