Research map / Time series / Causal and generative time series
High Dimensional Forecasting via Interpretable Vector Autoregression
William B. Nicholson, Ines Wilms, Jacob Bien, David S. Matteson
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use nested group lasso hierarchy instead of latent overlapping group lasso (LOG) (ablation) | HLag (component/own-other/elementwise) · Simulation Scenarios 1-5 (known max lag) | relative MSFE of LOG: 1.036 / 1.048 / 1.037 in Scenario 1; 0.946 in Scenario 3 | Improved |
| Use nested group lasso instead of LOG when maximal lag order is unknown (pmax=12) (ablation) | HLag · Simulation Scenarios 1-5 (unknown order) | relative MSFE of LOG: 1.138, 1.235, 1.094 | Improved |
| Use nested group lasso instead of LOG as pmax grows from 5 to 50 (ablation) | HLag componentwise · Robustness Scenario 5 | relative MSFE of LOG: 1.004 -> 1.114 | Improved |
| Use nested group lasso instead of LOG on lag order selection (L1 error) (ablation) | HLag elementwise · Scenario 1-5 | relative L1 lag error of LOG: 1.529 in Scenario 1; 1.058 in Scenario 5 | Improved |
| Use HLag hierarchical lag penalty instead of lasso as pmax increases | HLag vs Lasso VAR · Robustness Scenario 5 (pmax 1 to 50) | MSFE | Improved |
| Use HLag instead of lag-weighted lasso and Bayesian methods | HLag · Robustness Scenario 5 | MSFE | Improved |
| Use HLag instead of least squares AIC/BIC | HLag elementwise · Scenario 3 | MSFE: a factor of 10 | Improved |
| Use HLag instead of univariate AR benchmark | HLag · Scenario 1 | MSFE: more than twice as large | Improved |
| Add variables to VAR: loss in forecast accuracy for HLag versus Bayesian methods | HLag vs Bayesian VAR · Macroeconomic forecasting (Small-Medium to Large VAR) | forecast accuracy loss: 5% vs 46% | Improved |
| Use more flexible own-other/elementwise structure instead of componentwise as k grows (ablation) | HLag · Macroeconomic Large VAR (k=168) | wMSFE | Improved |
| Use HLag on small VAR (k=10) instead of univariate AR | HLag own-other · Macroeconomic Small-Medium VAR | wMSFE | No clear effect |
| Raise pmax from 4 to 13 (ablation) | HLag vs lasso · Macroeconomic VARs | wMSFE | Improved |
Closest papers by meaning
- Multivariate Time Series Forecasting needs Cross Variable Loss (NeurIPS 2026 · similarity 0.87)
- Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators (ICLR 2024 · similarity 0.86)
- Structured Matrix Basis for Multivariate Time Series Forecasting with Interpretable Dynamics (NeurIPS 2024 · similarity 0.86)
- An Analysis of Linear Time Series Forecasting Models (ICML 2024 · similarity 0.85)
- Linear Transformers as VAR Models: Aligning Autoregressive Attention Mechanisms with Autoregressive Forecasting (ICML 2025 · similarity 0.85)
- HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
