DistDF: A New Time-Series Forecasting Method Based on Distribution Alignment

Time-series forecasting is one of the most practical tasks in machine learning, from predicting stock prices to managing power grids. For decades, researchers have refined models by training them to minimize mean squared error. But the team behind DistDF, presented at ICLR 2026, asks an uncomfortable question: what if the training principle itself is fundamentally flawed?

Mean squared error, or MSE, is the standard that few researchers depart from. The logic is simple: the closer the predicted value is to the real one, the better the model. But the DistDF authors point to a fundamental flaw in this approach. MSE works point by point. It compares individual values and ignores how the data is distributed over time, how different moments in the series relate to each other, and what the structure of uncertainty looks like over the long term. A model trained on MSE can guess the next point with high accuracy and still completely miss hidden patterns: seasonality, volatility spikes, correlations between variables.

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

The choice of the Wasserstein distance is not accidental. Unlike other metrics, it accounts for the geometry of the data space and is sensitive to subtle structural differences between distributions. The joint version also captures dependencies among several variables at once, which is critical for multivariate time series, the kind that dominate real-world tasks. Energy consumption, commodity prices and network traffic are all systems in which variables are deeply interconnected, and those connections disappear under standard MSE training.

In practice, DistDF delivers convincing results. On standard time-series forecasting benchmarks, the new method outperforms its competitors, most noticeably at long forecast horizons, where error accumulation has traditionally been a critical problem. Notably, the improvement shows up not only in the accuracy of central predictions but also in the calibration of uncertainty: models trained through distribution alignment better recognize when they "don't know" and signal it more honestly through confidence intervals.

The practical implications of this work go well beyond academic interest. In finance, accurate estimation of tail risk is literally worth billions, and it is exactly where MSE is blindest that the Wasserstein distance sees best. Managing power grids requires predicting not only average demand but also extreme consumption peaks. Logistics and supply chains benefit from models that understand the structure of demand as a whole rather than as a set of independent points. In all these fields, moving from point forecasts to distributional forecasts means a qualitatively different kind of decision-making.

DistDF is a signal that the era of naive MSE minimization in forecasting is coming to an end. Distribution alignment as a training principle opens the way to models that do not merely memorize trends but genuinely understand the nature of temporal data. If the ICLR 2026 results are confirmed in production systems, we will see optimal transport theory, a mathematical discipline whose roots go back to the problems of Gaspard Monge in the 18th century, become a standard tool in the arsenal of modern data teams.