DistDF: a new method for time series forecasting via distribution alignment

Time series forecasting is one of the most applied tasks in machine learning: from predicting stock quotes to managing power grids. For decades, researchers have refined models by training them to minimize mean squared error. However, a team of scientists who presented the DistDF paper at ICLR 2026 poses an uncomfortable question: what if the learning principle itself is fundamentally flawed?

Mean squared error, or MSE, is a standard that few deviate from. The logic is simple: the closer the predicted value to the actual value, the better the model. But the DistDF authors point to a fundamental flaw in this approach. MSE works pointwise — it compares individual values, ignoring how data are distributed over time, how different moments of the series are connected, and what the structure of uncertainty looks like in the long term. A model trained on MSE can accurately predict the next point, but completely miss hidden patterns — seasonal patterns, volatility jumps, correlations between variables.

This is where DistDF proposes a different paradigm. Instead of comparing points, the method compares distributions. The key mathematical tool becomes the joint Wasserstein distance — a metric from optimal transport theory that measures how "expensive" it is to transform one probability distribution into another. In simpler terms: the model learns not just to guess a number, but to reproduce the entire character of data behavior — their variability, interdependencies, the shape of distribution tails. This is a fundamentally different level of understanding time series.

The choice of Wasserstein distance is not accidental. Unlike other metrics, it accounts for the geometry of data space and is sensitive to subtle structural differences between distributions. The joint version of this distance additionally captures dependencies between multiple variables simultaneously — which is critically important for multivariate time series, which predominate in real-world tasks. Energy consumption, commodity prices, network traffic — all of these are systems where variables are deeply interconnected, and these connections disappear under standard MSE training.

In practice, DistDF demonstrates convincing results. In standard benchmarks for time series forecasting, the new method outperforms competitors, especially noticeably on long-term prediction horizons, where error accumulation traditionally becomes a critical problem. Notably, quality improvement is observed not only in the accuracy of central predictions, but also in the calibration of uncertainty: models trained through distribution alignment better understand when they "don't know" — and signal this more honestly through confidence intervals.

The practical consequences of this work go far beyond academic interest. In the financial sector, the accuracy of tail risk assessment is literally worth billions — precisely where MSE is most blind, the Wasserstein distance is most sighted. Power grid management requires predicting not just average demand, but extreme consumption peaks. Logistics and supply chains benefit from models that understand demand structure holistically, rather than as a set of independent points. In all these domains, the transition from point forecasts to distributional forecasts means qualitatively different decision-making.

DistDF is a signal that the era of naive MSE minimization in forecasting is coming to an end. Distribution alignment as a learning principle opens the door to models that don't just memorize trends, but truly understand the nature of temporal data. If the results from ICLR 2026 are confirmed in industrial systems, we will witness how optimal transport theory — a mathematical discipline with roots in the problems of Gaspard Monge in the 18th century — becomes a standard tool in the arsenal of modern data teams.