Neural Conformal Control for Time Series Forecasting
Authors: Ruipu Li, Alexander Rodríguez
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Using real-world datasets from epidemics, electric demand, weather, and others, we empirically demonstrate significant improvements in coverage and probabilistic accuracy, and find that our method is the only one that combines good calibration with consistency in prediction intervals. We evaluate NCC on a variety of real-world datasets and against other state-of-the-art CP methods. |
| Researcher Affiliation | Academia | University of Michigan EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: Forward pass over NCC at time step T. Parameter: User-specified window size w, learning rate η, and a set of n decreasing target error rates A := α1, , αn. Input: Historical data, along with past non-conformative scores, coverage errors and quantiles, which we group and denote as D := {Xt, yt, st, {errα t , ˆqα t }α A}T t=1. Quantile predictor M. Output: Conformalized prediction intervals ˆCα T +1, α A. 1: For all α A, initialize qα 1 = 0. 2: Forward pass over quantile predictor: qα1 T +1, . . . , qαn T +1 = M(D) 3: for all i 1, . . . , n do 4: Calculate the average error rate for window errαi w (T) = 1 w PT t=T w+1 errαi t 5: Update qαi as qαi T +1 = qαi T + η( errα w(T) αi) 6: Conformalization step: ˆqαi T +1 = qαi T +1 + qαi T +1 7: if forward pass is used at test time (inference) then 8: TTA: Train learnable vector h Rn (until desired DCS) so that ˆqαi T +1 + hi monotonically increases as αi decreases. 9: end if 10: ˆCαi T +1 = {y|st(ˆy T +1, y) ˆqαi T +1} 11: end for |
| Open Source Code | Yes | Project page https://github.com/complex-ai-lab/ncc |
| Open Datasets | Yes | We evaluate on a diverse set of publicly available real-world datasets, including commonly-used datasets in time series forecasting (weather, electric) and datasets that demonstrate significant amount of temporal distribution shifts (flu, covid-19, smd). |
| Dataset Splits | Yes | For testing, we used COVID-19 hospitalization data from the 2021-2022 respiratory virus season within the covid-19 dataset, training on the data from the 10 weeks before the respiratory virus season. The test period spans 28 weeks. |
| Hardware Specification | No | The paper does not explicitly describe the hardware used for running its experiments. |
| Software Dependencies | No | The paper mentions software components like "seq2seq GRU model", "Theta model", "Informer model", "GRUs", and "GCN" but does not specify their version numbers. |
| Experiment Setup | Yes | Parameter: User-specified window size w, learning rate η, and a set of n decreasing target error rates A := α1, , αn. To make errα t differentiable, we approximate it as: errα t sigmoid st ˆqα t K where sigmoid(x) = 1 1+e x is the sigmoid function, and K is a constant. We evaluate on a diverse set of publicly available real-world datasets, including commonly-used datasets in time series forecasting (weather, electric) and datasets that demonstrate significant amount of temporal distribution shifts (flu, covid-19, smd). To illustrate that our method does not depend on a specific base forecaster, we get the point predictions using three base forecasters: seq2seq GRU model, Theta model (Assimakopoulos and Nikolopoulos 2000), and Informer model (Zhou et al. 2021). |