📊 statistics

Wasserstein Exponential Smoothing

This paper extends classical exponential smoothing to distributional time series within the Wasserstein space, offering a parsimonious model with a consistently estimated smoothing parameter that demonstrates practical effectiveness in forecasting high-frequency financial returns and household electricity demands.

Original authors: Takuo Matsubara, Peiwen Jiang, Minh-Ngoc Tran, Wilson Ye Chen

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Takuo Matsubara, Peiwen Jiang, Minh-Ngoc Tran, Wilson Ye Chen

Original paper licensed under CC BY 4.0 (http://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

Imagine you are trying to predict the weather. In the old days, you might have just looked at yesterday's temperature (a single number) and guessed today's would be similar, maybe a little warmer or cooler. This is like Exponential Smoothing (ES), a classic, simple, and very popular way to predict the future based on the past. It works by taking a "weighted average" of what you just saw and what you thought would happen, using a single dial (called a parameter) to decide how much to trust the new data versus the old trend.

But what if the thing you are trying to predict isn't just a single number? What if it's a whole shape?

The Problem: Predicting Shapes, Not Numbers

In the modern world, data often comes in the form of distributions (shapes).

Financial Example: Instead of just knowing the stock market went up 1%, you have a whole picture of how all the trades happened that day. Maybe the market was calm in the morning but chaotic in the afternoon. That's a "shape" of volatility.
Energy Example: Instead of knowing a house used 10 kWh of electricity, you have a distribution of when that power was used throughout the day. Did they use it all at once, or spread it out?

The problem is that you can't just add or subtract these "shapes" like you do with numbers. If you try to use standard math on them, it breaks. You need a new kind of geometry to measure the distance between two shapes.

The Solution: Wasserstein Exponential Smoothing (WES)

The authors of this paper created a new tool called Wasserstein Exponential Smoothing (WES). Here is how they did it, using simple analogies:

1. The "Rubber Sheet" Geometry (Wasserstein Space)
Imagine your probability distribution (the shape of your data) is drawn on a rubber sheet. To move from one shape to another, you don't just stretch or squash it randomly. You slide the rubber sheet along the smoothest, most efficient path possible. In math, this is called Optimal Transport or Wasserstein distance.

Analogy: Think of moving a pile of sand from one spot to another. The "Wasserstein distance" is the minimum amount of effort (energy) required to move every grain of sand to its new destination.

2. The New "Smoothing" Trick
In the old method (Exponential Smoothing), you take a step halfway between where you were and where you are now.

Old Way: New Guess = (1 - Dial) * Old Guess + (Dial) * New Data (This works for numbers).
New Way (WES): The authors realized that instead of adding numbers, you can "slide" the rubber sheet. They take the "Old Guess" shape and slide it along the smoothest path (the geodesic) toward the "New Data" shape. The "Dial" (smoothing parameter) decides how far you slide.
- If the dial is low, you barely move the old shape (you trust the past more).
- If the dial is high, you slide the shape almost all the way to the new data (you trust the new observation more).

3. The Magic of One Dimension
The paper focuses specifically on data that can be represented on a single line (like time or price). In this specific case, there is a mathematical shortcut: you can turn these complex shapes into simple lists of numbers (called quantile functions). This makes the "sliding" math easy to calculate, keeping the method just as simple and fast as the original old-school method.

How They Tested It

The authors didn't just do math on paper; they tested it in two real-world scenarios:

Stock Market Chaos: They looked at high-frequency stock trading data. Instead of predicting a single price, they predicted the distribution of daily returns.
- Result: WES was better at predicting the shape of the market's volatility than more complex, "fancy" models that tried to be too clever. It was the most accurate and stable.
Home Electricity: They looked at smart meter data from London homes. They predicted the distribution of electricity usage residuals (the weird, unpredictable parts of usage after removing normal patterns).
- Result: Again, WES beat the competition. It successfully tracked how the "shape" of energy usage changed from day to day.

The Bottom Line

The paper claims that Wasserstein Exponential Smoothing is a powerful, simple, and robust way to predict the future of "shape-based" data.

Why it's special: It takes a beloved, simple forecasting tool (Exponential Smoothing) and upgrades it to handle complex data distributions without losing its simplicity.
The Catch: Currently, it only works for data that fits on a single line (1D). It cannot yet handle complex, multi-dimensional shapes (like 3D clouds of data), because the mathematical "shortcut" they used doesn't work there yet.

In short, they found a way to make a classic, simple forecasting tool work for the complex, shape-heavy data of the modern world, proving that sometimes the simplest approach, when applied with the right geometry, is the most effective.

Technical Summary: Wasserstein Exponential Smoothing

Problem Statement

The analysis of distributional time series—where observations are probability distributions rather than scalar values—presents unique challenges distinct from classical time series forecasting. The state space of probability distributions is infinite-dimensional and lacks the vector-space structure required for standard operations like addition and scalar multiplication. Consequently, traditional forecasting methodologies, such as autoregressive (AR) models and exponential smoothing (ES), cannot be directly applied. While recent literature has adapted autoregressive frameworks to the Wasserstein space (e.g., Zhang et al., 2022; Zhu and Müller, 2023), the methodological toolkit remains underdeveloped compared to scalar settings. Specifically, there is a notable absence of a simple, robust, and parsimonious analogue to exponential smoothing for distributional data.

Methodology

The paper proposes Wasserstein Exponential Smoothing (WES), a generalization of classical ES to the space of probability distributions on $\mathbb{R}$ , denoted as $W_2(\mathbb{R})$ .

Core Framework

The methodology lifts the recursive interpolation of classical ES from Euclidean space to the Wasserstein space by replacing linear interpolation with geodesic interpolation under the Wasserstein distance.

State-Space Representation: The WES process models the evolution of a latent predictor distribution $\mu_t$ and an observable distribution $\nu_t$ . The observable is generated via a random pushforward map (noise) applied to the previous predictor.
Recursive Update: The predictor $\mu_t$ is updated as an intermediate point on the Wasserstein geodesic connecting the previous predictor $\mu_{t-1}$ and the current observation $\nu_t$ . Formally, for a smoothing parameter $\theta \in (0,1)$ :
$\mu_t = \left( (1-\theta) \cdot \text{Id} + \theta \cdot T_{\mu_{t-1} \to \nu_t} \right) \# \mu_{t-1}$
where $T_{\mu_{t-1} \to \nu_t}$ is the optimal transport map and $\#$ denotes the pushforward operation.
Quantile Representation: Leveraging the one-dimensional nature of the domain, the authors utilize the isometry between the 2-Wasserstein distance and the $L^2$ distance between quantile functions. This allows the complex distributional recursion to be expressed equivalently in terms of quantile functions $U_t$ and residuals $F_t$ :
$U_t(q) = U_{t-1}(q) + \theta F_t(q)$
where $F_t(q) = V_t(q) - U_{t-1}(q)$ represents the quantile residual. This formulation mirrors the error-correction form of classical ES.

Estimation

The smoothing parameter $\theta$ is estimated by minimizing the Minimum Wasserstein Prediction Error. Given a data sequence $\{\nu_t\}_{t=1}^T$ , the estimator $\hat{\theta}_T$ is defined as:
$\hat{\theta}_T = \arg \min_{\theta \in (0,1)} \frac{1}{T} \sum_{t=0}^{T-1} W_2^2(\mu_t^\theta, \nu_{t+1})$
This minimization is computationally feasible due to the tractable quantile representation in one dimension.

Key Contributions

Methodological Extension: The paper formulates a principled exponential smoothing procedure for distributional time series within the Wasserstein space, retaining the high parsimony of classical ES (governed by a single scalar parameter).
Estimation Theory: It proposes and analyzes a consistent estimator for the smoothing parameter based on minimizing the Wasserstein distance.
Theoretical Properties: The authors establish theoretical results including:
- Fréchet-Mean Stationarity: The WES process is shown to be stationary in terms of its Fréchet mean.
- Residual Autocovariance: The model-quantile residuals exhibit exponential decay in autocovariance.
- Consistency: The minimum Wasserstein estimator is proven to be consistent, converging to the true parameter as the sample size increases.
Empirical Validation: The method is applied to two distinct domains: high-frequency financial returns and household electricity demand residuals.

Results

Simulation Studies

Simulations using random shift and random sine pushforward maps demonstrate that the minimum Wasserstein estimator is unbiased and stable in finite samples. As the sample size increases, the variability of the estimator decreases, and the estimates center tightly around the true smoothing parameter.

Empirical Applications

The authors compared WES against four existing distributional time series benchmarks:

Wasserstein Autoregressive (WAR)
Autoregressive Transport Map (ATM)
Distributional Autoregressive Model (ODAM)
Distribution-on-Distribution Regression (DoDRM)

Financial Returns: Using daily distributions of 5-minute log returns for 10 major global equity indices, WES achieved the lowest Mean Wasserstein Prediction Error (MWPE) for all 10 indices. It was the unique member of the 90% Model Confidence Set (MCS) in 8 out of 10 cases and statistically outperformed all benchmarks in the majority of Diebold-Mariano tests.

Electricity Demand: Using daily residual demand distributions from 10 London households, WES again achieved the lowest MWPE for all 10 households. It remained in the 90% MCS for every household, whereas the best-performing benchmark (WAR) only appeared in 4 out of 10.

Significance and Claims

The paper claims that WES provides a compelling alternative to complex autoregressive formulations for distributional time series, particularly in applications where adaptivity and simplicity are paramount. The significance of the work lies in:

Bridging a Gap: It fills a critical methodological void by providing a distributional analogue to the widely used and robust exponential smoothing technique.
Computational Efficiency: By exploiting the one-dimensional Wasserstein geometry, the method avoids the computational burden of estimating autoregressive operators or transport maps, remaining analytically tractable.
Empirical Superiority: The results suggest that the parsimonious recursive structure of WES often yields superior predictive accuracy compared to more sophisticated autoregressive benchmarks in both financial and non-financial settings.

The authors note that the current framework is restricted to one-dimensional distributions due to the reliance on the exact isometry between Wasserstein distance and quantile $L^2$ distance. They acknowledge that extending this to multivariate settings ( $d > 1$ ) remains a non-trivial challenge for future research.