Conformal PID Control for Time Series Prediction
Authors: Anastasios Angelopoulos, Emmanuel Candes, Ryan J. Tibshirani
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on 4-week-ahead forecasting of statewide COVID-19 death counts in the U.S. show an improvement in coverage over the ensemble forecaster used in official CDC communications. We also run experiments on predicting electricity demand, market returns, and temperature using autoregressive, Theta, Prophet, and Transformer models. |
| Researcher Affiliation | Academia | Anastasios N. Angelopoulos University of California, Berkeley angelopoulos@berkeley.edu Emmanuel J. Candès Stanford University candes@stanford.edu Ryan J. Tibshirani University of California, Berkeley ryantibs@berkeley.edu |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. Algorithmic steps are described within the regular text using mathematical equations. |
| Open Source Code | Yes | We provide an extendable codebase for testing our methods and for the integration of new algorithms, data sets, and forecasting rules.1 1http://github.com/aangelopoulos/conformal-time-series |
| Open Datasets | Yes | COVID-19 death forecasting. ... The base forecaster ft that we use is the ensemble model from the COVID-19 Forecast Hub, which is the model used for official CDC communications on COVID-19 forecasting [10, 29]. Electricity demand forecasting. Next we consider a data set on electricity demand forecasting in New South Wales [18]. Return (log price) of Amazon, Google, and Microsoft stock [27]. Temperature in Delhi [37]. |
| Dataset Splits | No | The paper mentions retraining base forecasters and using a 'burn-in period' but does not specify explicit train/validation/test dataset splits with percentages, sample counts, or specific methods like k-fold cross-validation. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running its experiments. |
| Software Dependencies | No | The paper mentions 'darts' and 'sklearn.linear_model.QuantileRegressor' but does not specify version numbers for these software components. |
| Experiment Setup | Yes | In all cases except for the COVID-19 forecasting data set, we: re-train the base forecaster at each time point; construct prediction sets using the asymmetric (signed) residual score; and use a Theta model for the scorecaster. For the COVID-19 forecasting setting, we: use the given ensemble model as the base forecaster (no training at all); construct prediction sets using the asymmetric quantile score; and use an ℓ1-penalized quantile regression as the scorecaster... And lastly, in all cases, we use a tan function for the integrator with constants chosen heuristically, as described in Appendx C. With its default learning rate: η = 0.005 for ACI, and η = 0.1 ˆBt for quantile tracking... increases the learning rate for each method: η = 0.1 for ACI, and η = 0.5 ˆBt for the quantile tracker. |