Horizon-resolved Forecastability of Time Series via Auto Mutual Information
This paper demonstrates that auto-mutual information (AMI) serves as an effective pre-modelling diagnostic for assessing time series forecastability across specific horizons, revealing a systematic gradient where higher AMI values correlate with significantly lower forecasting errors compared to absolute autocorrelation.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of data, from stock markets to weather patterns, we often face a frustrating paradox: we can only know how well a prediction worked after we have made it. We build a model, run it into the future, and wait to see if the numbers match reality. But what if we could know, before writing a single line of code, whether a specific stream of numbers holds enough hidden order to be predicted at all? This question sits at the heart of a field called forecasting, where the goal is to turn history into a guide for the future. For decades, experts have relied on simple tools to measure how much a past event influences a future one, assuming that if the past and future are linked, a prediction is possible. However, these traditional tools often miss the subtle, complex ways data can be connected, especially when the relationship isn't a straight line. As businesses and governments increasingly rely on massive collections of data streams, the ability to quickly identify which streams are worth the effort to predict—and which are too chaotic to bother with—has become a critical skill.
A researcher named Peter Catt has proposed a new way to answer this question, offering a method to measure the "predictability" of a time series before any forecasting model is even built. Instead of guessing or waiting for results, his approach uses a concept called auto-mutual information. In plain terms, this measures how much knowing the present value of a number tells us about its value at a specific point in the future. If the present and future are tightly linked, the measure is high; if they are unrelated, it is low. Crucially, this method is designed to look at specific future timeframes, such as next month or next year, rather than giving a single, vague score for the entire dataset. The researcher tested this idea on a massive collection of nearly 48,000 monthly data series from a famous forecasting competition, covering everything from industry sales to demographic trends. The goal was to see if this pre-measurement could accurately rank which series would be easy or hard to predict.
The study found that this new measure works, revealing a clear and consistent pattern across thousands of different data streams. When the researcher sorted the data into groups based on their pre-measured predictability, a distinct gradient emerged. The series with the highest scores—those where the present and future were most strongly linked—were indeed the ones that forecasting models got right most often. Conversely, the series with the lowest scores were the ones that models struggled with. The difference was not trivial. For the models tested, the middle of the forecast errors for the most predictable group was nearly 30% to 37% lower than for the least predictable group. This suggests that the measure successfully identified a "richness" in the data that models could exploit. It is important to note that a high score does not guarantee a perfect prediction, nor does a low score mean a prediction is impossible; rather, it signals how much useful information is actually available in the history of that specific series to be used by a model.
One of the most significant findings was that this new method outperformed the standard tool used for decades: a simple measure of linear correlation. While the traditional tool looks for straight-line relationships between past and future, the new method captures more complex, curved, or irregular connections that the old tool misses. The advantage was small but consistent across all the different types of models tested, from simple statistical methods to advanced artificial intelligence. The study showed that this extra edge was most noticeable at longer time horizons, where the simple linear tool often failed to see any connection at all, while the new method still detected a signal. This confirms that the extra complexity of the new measure is justified, as it uncovers patterns that simpler tools overlook.
The research also clarified what this measure actually tells us about the nature of prediction. It distinguishes between the quality of the data itself and the skill of the model used to analyze it. A series might have a high score, meaning it contains a lot of predictable structure, but if a model is too simple to find that structure, the prediction will still fail. Conversely, a series with a low score might be hard to predict simply because the information is weak, regardless of how smart the model is. By measuring the data first, this approach allows analysts to decide where to invest their effort. It acts as a filter, identifying which data streams are likely to yield good results and which are likely to remain chaotic, saving time and resources. The study validated these findings using synthetic data where the true answer was known, confirming that the method accurately detects non-linear connections that other tools miss.
Ultimately, this work shifts the focus from a reactive process to a proactive one. Instead of building a model and then hoping for the best, analysts can now look at the data and ask, "Is there anything here to find?" The measure is not a crystal ball that predicts the future, but a diagnostic tool that reveals the potential for prediction. It works by scanning the history of a series to see how much of the future is encoded in the present. While the method requires recalibration when applied to new sets of data, it offers a powerful, model-free way to understand the limits of what can be known. In a world drowning in data, the ability to distinguish between a signal and noise before the work begins is a valuable step forward, turning the art of forecasting into a more precise science.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.