How much are prices rising?
X is supposed to predict inflation in the future
September 24, 2023, 2:58 p.m
Listen to article
This audio version was artificially generated. More info | Send feedback
Can machine learning help measure inflation expectations? Yes, say researchers at the Frankfurt School of Finance & Management. Your data basis: posts on social media.
The Frankfurt School of Finance & Management has developed an inflation expectations index based on social media posts from the Twitter platform, which was recently renamed X. You are the leader of the project. How did you go about it?
Nora Lamersdorf: The data source for our index was around ten million German tweets posted since 2011. First we downloaded these via a so-called API. This is a programming interface that can be used to unlock additional functions. We selected potentially relevant tweets using specific keywords. These are words like “prices”, “expensive”, “cheap” or “inflation”. Our index results from evaluating these tweets using machine learning methods.
That sounds very simple at first. But how accurate can an inflation index based on a platform as chaotic and polemical as Twitter be?
In fact, during the evaluation period, the index correlates strongly with the realized inflation rates and inflation expectations of households, which come, for example, from survey data from the Bundesbank. But it is only so accurate because we have cleaned the data source beforehand. For example, to clean it of bot activity and other interference signals that could distort the result, the index uses the possibilities of machine learning.
How does this work?
We divided the text corpus into different topics and filtered out those tweets that addressed inflation topics. These tweets are classified as “rising,” “falling,” or “neutral” by a pre-trained language model. To generate enough training data for this fine-tuning process, the index also uses OpenAI’s ChatGPT API. Our trained model then aggregates all tweets classified as rising or falling for each day into a daily inflation index. In addition to a German inflation expectations index, the model can also reflect regional indices.
Do you weight tweets based on relevance? If in doubt, an economics professor knows more about inflation than a tennis teacher…
No, we are not weighing this up yet. I don’t think that would make scientific sense either. It’s about the general inflation expectations of the population and every assessment is initially valid. The expectation is directly related to actual inflation. It’s really about each individual’s expectations because this also affects each individual’s consumption decisions. Of course, there are experts whose status could shape opinions a little. But I would also say that there are groups that may not necessarily be reached by these experts, and who may even follow people who rail against such experts or the government.
Do you recognize that your index reacts to current events? Such as interest rate decisions by central banks.
The index provides additional explanatory value and improved forecast accuracy compared to existing measurement methods of quantitative inflation expectations. However, the big advantage of such an index is its daily availability – in contrast to survey data, which only appears once a month. Among other things, we can show that the index reacts to monetary policy measures within a few days; For example, it rises when restrictions are relaxed and falls after unexpected tightening. This effect is particularly noticeable in tweets from private users in the current phase of increased inflation. Overall, the new index is a valuable tool for market actors and policy makers to identify prevailing inflation expectations in a timely manner.
If models like your index prove successful, doesn’t such technology also allow for misuse?
It is true that big data models could theoretically be used in ways that can raise moral questions. But there are definitely a wide range of topics where such data is useful – both positive and negative. Take, for example, the banking crisis in the USA. This case highlighted how quickly information spread and how well-connected investors were able to withdraw their money very quickly. In such a situation, Twitter data could help to identify such developments at an early stage.
What else could be predicted?
In principle, you could analyze Twitter data on almost any topic imaginable to find out whether people are talking about certain topics, the tone in which they do so, and whether they might be concerned. Similar analyzes could also be carried out on other social media to identify trends early and potentially mitigate negative impacts. But of course we have to respect privacy and data protection. Data should only be used aggregated and anonymized. For example, you shouldn’t focus on individual users.
Leon Berent spoke to Nora Lamersdorf