Abstract:We present data-dependent learning bounds for the general scenario of non-stationary non-mixing stochastic processes. Our learning guarantees are expressed in terms of a data-dependent measure of sequential complexity and a discrepancy measure that can be estimated from data under some mild assumptions. We also also provide novel analysis of stable time series forecasting algorithm using this new notion of discrepancy that we introduce. We use our learning bounds to devise new algorithms for non-stationary time series forecasting for which we report some preliminary experimental results.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to solve the time - series prediction problem of non - stationary, non - mixing stochastic processes. Specifically, the goals of the paper include: 1. **Provide learning bounds applicable to non - stationary, non - mixing processes**: - The paper proposes data - dependent learning bounds for non - stationary, non - mixing stochastic processes. These bounds are measured by path - dependent generalization errors and do not rely on stationarity and mixing assumptions. 2. **Introduce new metric and complexity concepts**: - Two key concepts are introduced: sequential complexity and discrepancy measure between distributions. These concepts are used to express and derive new learning bounds. - In particular, a metric method based on the discrepancy between sample and target distributions is proposed, which can be estimated from data under some mild assumptions. 3. **Analyze stable time - series prediction algorithms**: - Using the newly proposed discrepancy measure, stable time - series prediction algorithms are analyzed, extending previous results in the stationary mixing scenario. 4. **Develop new time - series prediction algorithms**: - Based on the above theoretical results, a new non - stationary time - series prediction algorithm is designed, and some preliminary experimental results are reported. 5. **Cover a wide range of model types**: - The proposed method is applicable not only to bounded - memory models but also to unbounded - memory models, covering most methods in practical applications, such as autoregressive models and state - space models. ### Summary The core contribution of the paper lies in providing a new theoretical framework and algorithm for time - series prediction of non - stationary, non - mixing stochastic processes. By introducing new metric tools and complexity concepts, the author is able to provide more widely applicable learning bounds for such complex processes and develop new prediction algorithms on this basis. This helps to overcome the strong assumption limitations of stationarity and mixing in traditional time - series analysis.

Theory and Algorithms for Forecasting Time Series

Globally Optimal And Adaptive Short-Term Forecast of Locally Stationary Time Series And A Test for Its Stability

Probabilistic Learning of Multivariate Time Series with Temporal Irregularity

Learning Predictive Leading Indicators for Forecasting Time Series Systems with Unknown Clusters of Forecast Tasks

Learning Graph Structures and Uncertainty for Accurate and Calibrated Time-series Forecasting

Learning Leading Indicators for Time Series Predictions

Foundations of Sequence-to-Sequence Modeling for Time Series

Nonparametric risk bounds for time-series forecasting

Consistent Long-Term Forecasting of Ergodic Dynamical Systems

Algorithmic Information Forecastability

Development of a Time Series Forecasting Method Based on Entropy Measures of Diagnostic Information

Machine Learning Algorithms for Time Series Analysis and Forecasting

NonSTOP: A NonSTationary Online Prediction Method for Time Series

Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling

Finite-Time Error Bounds For Linear Stochastic Approximation and TD Learning

Equivariant online predictions of non-stationary time series

Adaptive time series forecasting with markovian variance switching

Empirical Risk Minimization for Time Series: Nonparametric Performance Bounds for Prediction

Fast rates in learning with dependent observations

Discovering Predictable Latent Factors for Time Series Forecasting

Calibrated Probabilistic Forecasts for Arbitrary Sequences