Introduction
In the era of big data, the prevailing wisdom suggests that larger datasets lead to more robust models. However, time series forecasting presents a unique challenge: non-stationarity. Real-world events—such as the COVID-19 pandemic, policy shifts, or sudden market crashes—create structural breaks or concept drifts that render historical data obsolete. For instance, predicting tourism trends in 2024 using data from the 1980s may introduce significant noise, as the underlying dynamics of travel have fundamentally shifted.
While sophisticated models like Long Short-Term Memory (LSTM) networks or Exponential Smoothing (ETS) naturally prioritize recent observations, they still process the entire historical record, which can degrade performance. This paper introduces OSP-TSP (Optimal Starting Point Time Series Forecast), a framework designed to automatically identify the most relevant historical starting point for training, thereby truncating irrelevant past data to minimize future prediction errors.
Methodology: The OSP-TSP Framework
The OSP-TSP approach utilizes a two-step meta-learning process to move away from subjective or manual data truncation.
1. The Meta-Learning Step
Instead of heuristic-based selection, the authors train a supervised machine learning model (using XGBoost or LightGBM) to predict the optimal starting interval. To make this computationally feasible, the time series is divided into $m$ sub-intervals rather than searching for a specific point. The meta-model learns to identify which interval contains the starting point that minimizes forecast error.
A critical component of this step is Feature Extraction. The framework employs the tsfeatures R package to convert raw time series into descriptive statistical vectors. These features capture essential characteristics such as trend strength, seasonality, linearity, and autocorrelation.
2. The Forecasting Step
Once the optimal starting point (OSP) is identified, the series is truncated. Standard "base" forecasting models—including ETS, thetaf, ARIMA, and Neural Network Autoregression (nnetar)—are then applied to the shortened series. This ensures the model focuses exclusively on the most relevant data regime.
Data and Experimental Setup
The study's efficacy was validated using a massive and diverse collection of datasets:
- M4 Dataset: A benchmark containing 100,000 time series across various frequencies (yearly to hourly) and domains (finance, macro, micro, etc.).
- Real-World Indicators: Five specific economic datasets including GDP and Exchange Rates.
- Simulated Data: Synthetic series generated via the GRATIS method to test the generalizability of models pre-trained on artificial data.
Key Insights
The research highlights several critical findings:
- The "Less is More" Paradox: Using a subset of recent data consistently outperforms using the full historical record, particularly in the presence of structural shifts.
- Feature Importance: Curvature and Linearity emerged as the most pivotal features for determining the OSP in seasonal models. For high-frequency data, autocorrelation features after differencing were the primary drivers.
- Transferability: OSP models pre-trained on the M4 dataset or simulated data effectively generalize to entirely different real-world datasets, reducing the need for task-specific training.
- Interval Averaging: The authors found that averaging forecasts from several points within the predicted "optimal interval" provided greater stability and accuracy than picking a single point.
Performance Results
The OSP-TSP approach demonstrated consistent improvements across multiple data frequencies. In the M4 dataset, the method significantly reduced the Mean Absolute Scaled Error (MASE) compared to models trained on the total series.
As shown in the results, for yearly data using an ETS base model, the MASE dropped from 3.44 (total series) to 2.51 (OSP-TSP). The framework proved to be model-agnostic, improving performance for both simple statistical models and complex neural networks. Notably, OSP-TSP outperformed traditional CUSUM-based change-point detection because it optimizes specifically for forecast error rather than just statistical shifts in mean or variance.
Conclusion and Future Work
While OSP-TSP offers a powerful solution for non-stationary forecasting, it is computationally demanding. Training meta-models for 100,000 series remains a hurdle for real-time applications. Future research will explore parallel computing to enhance speed and investigate how OSP benefits scale with even more complex architectures. Additionally, refining the selection heuristics for the "best" interval remains a key area for improvement.
Ultimately, OSP-TSP challenges the "more data is better" mantra, providing a data-driven path to more accurate and relevant time series forecasting.
Summary updated