Dec 15, 2021
FORECASTING INFLATION THE DATA SCIENCE WAY
Inflation has been an imperative driver of macroeconomic markets recently and a hot topic of conversation among market participants. Governments and central banks around the world are deploying trillions of dollars worth of fiscal and monetary stimulus to battle the economic effects of the pandemic. This stimulant activity, in collaboration with the gradual reopening of economies (which have for a fact created increased demand for goods and services), has piqued market participants’ focus on inflation. Some believe that the high inflation prints are transitory, while others think the price increases could persist for a longer period of time.
Understanding and forecasting future inflation rates is an extremely pivotal activity to various stakeholders including policymakers and investors. The textbook definition of Inflation says that it is the change in prices within an economy over time. We often consider rising prices to be a bad thing, but when moderated these price increases are indicative of a strong and growing economy.
Inflation expectations have proven to be channels whereby monetary policy influences the economic activity. They play a salient role in the process of consumer goods price formation at different producer and retailer levels.
Developing a robust forecast of inflation expectations has become a requisite for businesses, financial institutions, policy makers, and even individuals to make more informed decisions in the marketplace.
Challenges of Forecasting Inflation
Data scientists often gather data to the hilt to extract new insights about the problems or questions at hand. Taking a data-science approach to forecasting inflation potentially involves a combination of many forecasts, perhaps with various conviction levels, to form some new linking opinions. Even if this sounds fairly straightforward, it does come with several not so simple to conquer challenges. We plan on firstly summarizing these challenges and then suggesting ways to seek to alleviate some of its effects.
One of the most eminent challenges would be defining the term inflation itself. It is notably not easy to concise it into a single term since the process has many variations associated to it. Apart from that, the rapidly increasing volume of data fabricates the issue of Data Proliferation. This makes the data comprehension task extremely gruelling and demanding. Moreover, specifications for various products and their proficiency change over a given time period hence making it difficult to find the historical comparability for the same.
In addition to that, forecast horizon plays an exceptionally crucial role in deciding the time frame for which further forecasts are to be prepared. When it comes to applying it to inflation, it becomes a tedious task since estimates for inflation can vary in terms of their forecast horizons. The fore-mentioned challenges are the most prominent challenges of the process.
Considering all these challenges it might seem that coming up with a well-rounded and accurate approach to measure all aspects of inflation is a fairly tedious task but that’s the farthest from reality. There’s a diverse range of methods being applied including Univariate KNN, KNN with explanatory variables and Random Forests. These methods use a variety of packages and algorithms to come up with the desired insights, however, one of the most interesting methods that has recently been in discussion is the Convolutional Modelling of Time Series Data.
For this analysis, it is necessary to develop a Convolutional Neural Network (CNN) to predict inflation for a specified time period say 12 months.
Convolutional Neural networks are basically considered to be a series of deep learning algorithms which were originally designed to understand the classification of images. The network takes an image and basically passes that image through a set of filters applying weights to different aspects of the image and ultimately provides a prediction. This works as a feature engineering system whereby, over time, the network “learns” what aspect filters are most important in classifying an image.
A similar method can be applied to time series the only difference being that Time Series data would not consist of the same features as that of an image but it would be consisting of dimensional features instead. Moving on, this method uses a one-step rolling prediction that means that for every month in the test set the inflation rate for each of the next twelve months is predicted and then the observed values of inflation are used to predict the next set of twelve months. This is a fairly realistic approach, as in forecasting the next period of inflation the analyst will have observed rates of all prior periods.
Conclusion:
To wrap up, there are more specific challenges to overcome when analysing forward-looking inflation, which turns out to be a key component of asset allocation decisions. In this blog we tried covering the same while discussing the current trends and methods being applied to counter those challenges and providing the market participant with an overview of the whole process.
References:
https://www.sciencedirect.com/science/article/pii/S2666143820300120
https://www.kaggle.com/donatasbag/eu-inflation-forecasting-using-machine-learning
https://towardsdatascience.com/deep-learning-to-predict-us-inflation-70f26405bf76
https://www.twosigma.com/articles/forecasting-inflation-like-a-data-scientist-2021-edition/
