Research map / Time series / Long-term forecasting

Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters

NeurIPS 2026 · Time series › Long-term forecasting · approach: Attention for time series forecasting · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Large-scale empirical study (4,218 runs) of synthetic time series augmentation with a difficulty-conditioned bundle generator across five forecasting architectures and seven benchmarks, finding effects depend strongly on architecture.

Problem

Whether and when synthetic time series data helps training deep learning forecasters, given that its usefulness is poorly understood compared with LLMs and that augmentation can degrade performance under distribution mismatch.

Approach

A controllable synthetic generator with four difficulty-conditioned bundle types (Seasonal-Trend, Non-stationary Regime, Long Memory, Volatility Events), Dirichlet variance allocation, cross-channel correlation, and mixing/curriculum strategies, used to systematically sweep augmentation factors across architectures and datasets.

Key results

• Augmentation hurts in 67% of Group 3 trials; TimesNet improves in 67% and iTransformer in 71% of trials while DLinear improves in only 7%.
• TimesNet trained on 10% of Weather with synthetic augmentation reaches MSE 0.2841 vs 0.2850 for full-data real-only baseline (H=336).
• Hard curriculum switching increases MSE by 23.6% over static mixing; Seasonal-Trend bundle is least harmful (mean degradation 10.5%).

Surprising findings

• Hard-switch curriculum annealing is catastrophic, degrading MSE by about 24% relative to static mixing.
• Synthetic augmentation with TimesNet on 10% of Weather surpasses full-data real-only training in one setting.

Limitations

• Only trend-dominated long-term forecasting benchmarks were tested; finance, healthcare and regime-shift datasets are absent.
• Models are trained from random weights for 10 epochs; foundation-model pre-training or fine-tuning was not studied.
• Evaluation is limited to point forecasting (MSE and MAE).

Experimental setup

Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic synthetic bundles (ST, NR, LM, VE)

Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic

Metrics: MSE MAE

Baselines: DLinear SegRNN TimesNet iTransformer PatchTST

Builds on: TSLib

Closest papers by meaning

• Are Language Models Actually Useful for Time Series Forecasting? (NeurIPS 2024 · similarity 0.88)
• A Closer Look at Transformers for Time Series Forecasting: Understanding Why They Work and Where They Struggle (ICML 2025 · similarity 0.87)
• Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.87)
• TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting (ICLR 2024 · similarity 0.87)
• SynTSBench: Rethinking Temporal Pattern Learning in Deep Learning Models for Time Series (NeurIPS 2025 · similarity 0.86)
• Timer: Generative Pre-trained Transformers Are Large Time Series Models (ICML 2024 · similarity 0.86)