TL;DR
This paper introduces Continuous Horizon (CH) forecasting, a new paradigm that treats future trajectories as continuous functions rather than discrete vectors. By redefining the forecast horizon as a queryable variable, CH achieves lower estimation variance, enables $O(1)$ targeted horizon querying, and naturally supports interpolation and extrapolation to unseen time points.
What Problem Does This Paper Address?
For decades, multi-horizon forecasting has been locked into three discrete strategies:
- Recursive (REC): Predicting one step at a time and feeding it back. (Suffers from error accumulation).
- Direct (DIR): Training a separate model for every future step. (Statistically inefficient and high variance).
- Multi-Step (MS): Predicting a fixed-length vector of future values all at once. (Rigid output grid, unable to query arbitrary time points).
The authors argue that this discrete view is a "historical artifact." Real-world phenomena are inherently continuous. Existing methods struggle to predict "between" the grid points (interpolation) or "beyond" the trained window (extrapolation) without clunky workarounds. Furthermore, they lack a unified theoretical framework to explain why some strategies perform better than others in different noise regimes.
Methods
The core innovation is the Continuous Horizon (CH) strategy. Instead of a model $g$ outputting a vector, it outputs a continuous operator: $\hat{y}_{t+h} = g(h \mid \mathbf{y}_{:t})$.
To implement this, the authors use a bilinear parametrization (similar to "branch-trunk" architectures like DeepONet): $$g(h \mid \mathbf{y}_{:t}) = \sum_{k=1}^K \theta_k(\mathbf{y}_{:t}) \phi_k(h)$$ Here, $\theta_k$ are coefficients generated by an encoder (the "branch") based on historical data, and $\phi_k(h)$ are basis functions (the "trunk") that depend only on the horizon $h$. These basis functions can be learned embeddings, Fourier series, or polynomials.
The paper provides two major theoretical pillars:
- Universal Operator Approximation Theorem: Proving that this bilinear form can approximate any continuous forecast function.
- Continuous Bias-Variance Decomposition: A formal proof showing that CH leverages "continuous effective horizon information" to achieve strictly lower estimation variance than discrete MS or DIR strategies.
Datasets
The authors validated CH across a diverse set of benchmarks:
- Simulated Autoregressive Processes: Linear (LAR), Nonlinear (NLAR), and Smooth-Transition (STAR) models to test bias-variance behavior.
- Gaussian Processes (GP): Five different regimes (mean-reverting, periodic, etc.) to evaluate interpolation and extrapolation capabilities.
- M-Series Benchmark: 16 large-scale real-world datasets from the famous M1, M3, M4, and Tourism competitions, totaling over 100,000 time series.
Key Insights and Ideas
- Discrete is just a special case: The authors show that REC, DIR, and MS are actually just coarse discretizations of the CH operator.
- The "Variance Dominance" Effect: Because CH treats the horizon as a continuum, it effectively uses more "information" from the training data to regularize the forecast. This leads to a smoother, more stable fit, especially in high-noise scenarios.
- Targeted Querying: Because the basis functions $\phi_k(h)$ can be pre-cached, querying a specific future time point $h$ (even a fractional one like $h=2.5$) becomes a simple $O(1)$ dot product.
- Aliasing Avoidance: Discrete models can suffer from aliasing if their output grid is too sparse to capture high-frequency oscillations. CH avoids this by learning the underlying continuous signal.
Key Results
- Accuracy: On the M-series benchmark, CH outperformed the standard Multi-Step strategy in 40 out of 64 encoder/frequency pairs, with a mean improvement of 2.6%.
- Long Horizons: The advantage of CH grew with the horizon length. For M4 Hourly data ($H=48$), CH reduced error by up to 49% compared to MS when using a Transformer encoder.
- Robustness: In high-noise simulations, CH showed 40% to 55% lower variance than the Direct strategy.
- Zero-Shot Capabilities: CH successfully extrapolated to horizons twice as long as those seen during training, significantly outperforming the recurrent rollouts of REC models.
Limitations and Future Work
- Point Forecasts: The current work focuses on point estimates. Extending this to probabilistic forecasting (e.g., predicting continuous density functions over time) is a logical next step.
- Foundation Models: Most current "Time Series Foundation Models" still use fixed output grids. Integrating CH into these large-scale models could make them much more flexible.
- Basis Constraints: The authors suggest that future work could explore constraining the learned basis functions (e.g., enforcing orthonormality) to improve interpretability and stability.
- Very Long Horizons: For extremely long-term forecasting, hybrid strategies combining CH with recursive elements might be necessary to manage computational costs.
Summary updated