Time Series Forcasting
- Citation Author(s):
-
Longfei Liu
- Submitted by:
- Dan Wu
- Date Created:
- Last updated:
- DOI:
- 10.21227/5asj-pf08
- Links:
157 views
Abstract
The Time Series Library (TSL) bundles widely used long-horizon forecasting benchmarks spanning diverse domains—energy (ETTh1/ETTh2/ETTm1/ETTm2, Electricity/ECL), transportation (Traffic), climate (Weather), macro-finance (Exchange), and public health (Illness/ILI). These datasets cover granularities from 15-minute to daily, with input windows typically 96–720 steps and forecast horizons matching 96/192/336/720. They feature real-world challenges such as non-stationarity, seasonal-trend interactions, regime shifts, missing values, and sensor noise. Standard splits (train/val/test) and metrics (MSE/MAE) enable consistent comparison across models, while cross-dataset and cross-domain transfers support evaluation of robustness and generalization.
Instructions:
We adopt the Time Series Library to evaluate long-term forecasting methods under realistic, heterogeneous conditions. The collection’s breadth—covering industrial electricity loads, road traffic occupancy, multi-variable meteorology, currency rates, and influenza incidence—allows us to test whether a model scales across sampling rates, variable counts, and shift types. Following common practice, we use fixed input windows and multi-step outputs, report MSE/MAE, and include zero-shot or domain-shift protocols (e.g., ETTh1→ETTh2) to probe generalization. This setting stresses models to capture multi-scale periodicity, preserve trend coherence, and remain stable under distribution drift—requirements central to practical long-horizon forecasting.
