TL;DR
This paper introduces a novel nonparametric approach for predicting quantiles of time series, outperforming traditional methods on real-world call center data. The method, called QuantileExpertMixture, combines multiple "expert" predictors using an adaptive weighting scheme, offering robust and accurate quantile forecasts without strong assumptions about the underlying data distribution.
What Problem Does This Paper Address?
Forecasting future values of time series is crucial across many domains, from finance to genetics. While predicting the mean of a time series has been extensively studied, many applications require predicting quantiles. This is particularly important in risk management (e.g., Value at Risk in finance) or resource planning (e.g., staffing levels in call centers), where understanding the range and probability of outcomes is more critical than just the average. Existing methods often rely on strong assumptions about the data's statistical properties (e.g., stationarity, boundedness, or specific distributional forms), which may not hold in practice. This paper aims to address the problem of sequential quantile prediction for real-valued time series under minimal assumptions, offering a more flexible and robust alternative.
Methods
The core of the proposed method, QuantileExpertMixture, is an "expert aggregation" strategy. This involves:
- Elementary Experts: A set of "experts" are defined, each being a nearest-neighbor (NN) based predictor. These experts scan past observations to identify patterns and make predictions. Specifically, an expert $h_{k,l}^{(k,l)}$ looks at the $l$ nearest neighbors of the most recent observation within a window of $k$ past observations.
- Pinball Loss Minimization: The prediction task is framed as minimizing the "pinball loss" function, which is the standard loss function for quantile estimation. This is a key departure from methods focused solely on squared error.
- Adaptive Weighting: The predictions from individual experts are combined using a weighted average. The weights are dynamically updated based on the past performance of each expert, giving more influence to those that have been more accurate in predicting quantiles. This is achieved through an exponential weighting scheme that penalizes poor performance.
- Nonparametric Approach: The method does not assume the time series follows a specific parametric model (like ARMA) or is bounded. It relies on the data itself to learn the patterns, making it more adaptable to diverse real-world scenarios.
Datasets
The paper evaluates its method on 21 real-world time series representing daily call volumes entering call centers. These series have an average of 760 data points. The authors also define a set of 91 selected dates for evaluating forecasts. The data is available at http://www.lsta.upmc.fr/doct/patra/. No new benchmark datasets were introduced; existing real-world data was used for evaluation.
Key Insights and Ideas
- Quantile Prediction is Crucial Beyond Mean: The paper strongly emphasizes that for many practical problems (especially in finance and resource management), predicting quantiles is more informative and actionable than predicting the mean.
- Robustness through Minimal Assumptions: By framing the problem as nonparametric quantile prediction and using an expert aggregation strategy, the method achieves consistency without requiring strict statistical assumptions about the time series, making it broadly applicable.
- Expert Aggregation for Adaptability: The core idea of combining multiple specialized predictors (experts) and adaptively weighting them based on performance is a powerful paradigm for handling complex, evolving time series patterns.
- Pinball Loss as the Right Tool: Explicitly using the pinball loss function for quantile prediction, rather than adapting squared-error methods, is key to the method's effectiveness in quantile forecasting.
Key Results
The QuantileExpertMixture strategy demonstrated strong performance on the call center data:
- Superior Performance: In terms of Average Absolute Error (which is proportional to pinball loss for $\tau=0.5$), the QuantileExpertMixture strategy outperformed all other methods tested, including simple moving average (MA), autoregressive models (AR, QAR), day-of-the-week models, and Holt-Winters.
- Competitive Results: For $\tau=0.1$ and $\tau=0.9$, the method showed comparable or better performance than the autoregressive quantile model (QAR(7)). For $\tau=0.5$, it significantly outperformed QAR(7).
- Consistency: Theoretical results show that the proposed strategy is consistent, meaning it asymptotically achieves the best possible quantile prediction loss for a wide class of stationary and ergodic time series.
Limitations and Future Work
- Computational Complexity: While the theoretical framework is elegant, the practical implementation involves selecting a finite grid of experts (k, l), which can still be computationally intensive. The paper mentions using a specific grid size for experiments.
- Choice of Parameters: The performance can depend on the choice of learning parameter $\eta_n$ and the truncation parameter $\delta$. The paper suggests $\eta_n = 1/\sqrt{n}$ for theoretical consistency and discusses practical considerations for parameter tuning.
- Handling Ties: The theoretical results rely on the assumption that ties in nearest neighbor searches occur with probability zero. The paper acknowledges that dealing with ties might require further investigation, referencing existing work in this area.
- Future Work: The paper focuses on the theoretical consistency and experimental evaluation of the proposed method. Further research could explore more efficient ways to select experts, optimize parameter tuning, and extend the method to multivariate time series or different types of quantile prediction problems.
Summary updated