Tracking inflation news

The most important step in any Machine Learning project is certainly to establish the use case. Inflation offers a perfect one in macro finance.

Can't ignore inflation

Inflation is key to the central banks outlook and, therefore, interest rates expectations. The current low inflation regime allows central banks to keep supporting the economy, while governments and firms borrow cheaply. Bring back some inflation, for whatever reason, and it would be a game-changer for strategic asset allocation.

Amid the global pandemic, the inflation outlook is subject to unusual and powerful opposing forces: a depressed global demand exerts downside pressure on prices, but profound disturbance and changes to supply chains could push costs higher, at least in a few key sectors.

Machine Learning use case

Inflation is a particularly difficult economic variable to forecast.

To build time-series models is only going to be one step in the inflation forecast process. It for sure won't be enough.

What makes a good inflation forecast over a few months horizon (say 1 to 6 months) is the inclusion of all sorts of micro news which can have a meaningful impact on inflation volatility. Often, there is no pre-established time series for these events, which makes accurately forecasting inflation challenging.

Just to give a few examples of such "micro news":

  • indirect taxes changes
  • unusual sales timing
  • weather effects on food prices
  • postal services price changes
  • supermarkets price wars
  • airline competition
  • regulated price changes
  • methodological changes

The list could certainly go on for pages. All these types of events can have a meaningful impact and they won't be easy to monitor through usual or alternative data. Rather, they appear first as "qualitative news".

To just keep track of these events is difficult, even more so when one deals with several countries and languages. To stay informed really is a permanent struggle and that is before we start analysing the implications of these specific news.

That is exactly where unstructured data models can help.

Introducing the Inflation NewsBot

To respond to these challenges, we have trained an Inflation NewsBot. It does this one thing: to warn something relevant to the short-term inflation outlook is happening, so the analyst or the risk taker can focus on... analyzing what the news mean for the inflation forecast and the markets.

Once the relevant news detection system is in place, quantitative analysis becomes possible again. But the process inevitably starts with selecting the relevant piece of information.

In short, the algo performs the following:

  1. a simpe search, downloading articles which have some connection with inflation, from 50,000 or so potential sources
  2. filtering, in which articles the closest to a number of inflation-related situation are selected using a pre-trained language model
  3. classification, which relies on a language model specially trained to distinguish between relevant and non-relevant inflation articles
  4. analysis, during which several language models summarize the information, sort the news by location and theme, information which can then be quantified and used as an input to further modelling tasks.

To teach the classification model what is relevant and what is not, we have manually classified more than 20,000 inflation-related news, in several languages, published on news websites over the last three years. With the eyes of inflation forecasters, we have distinguished between non-relevant and relevant news.

Then we have fine-tuned a state-of-the-art neural network language model to perform the classification task to decide whether an article is relevant or not for the inflation outlook.

The results are very encouraging by the literature standards (you can get some more details here). In essence, the model is only sightly less accurate than a trained economist. But it is incredibly faster and can process a lot more information.

To conclude, the use case is the key ingredient to a successful ML project, in finance and anywhere else. There has to be something the machine does significantly better than a human-being. In this case, it is the quantity aspect: there is no way even a team of analysts could process as much information as the NewsBot.