Selected work

Representative Papers

Scale-Invariant Training for Time Series Foundation Models

Oct 05, 2026

This study addresses the issue of "scale contamination" in time series foundation model training, where scale inversion causes high-magnitude samples to dominate gradients. To overcome this, we propose a scale-invariant training method employing Reversible Instance Normalization (ReVIN). For homogeneous loss functions such as mean squared error, the loss is computed directly on scaled targets, ensuring scale-independent optimization trajectories. This work provides the first theoretical proof and correction of biases in existing normalization schemes, establishing a unified loss paradigm that requires only a single line of code for integration. Experiments demonstrate that our approach consistently reduces the MASE metric across four architectures, yielding average improvements of approximately 20% on GIFT-Eval and M-competitions, thereby significantly enhancing forecasting accuracy and robustness.

0 citationsRead paper

The Nixtlaverse: An Open-Source Ecosystem for Forecasting

Sep 30, 2026

This study addresses the incompatibility of data interfaces among statistical, machine learning, and neural network model families in forecasting software, which necessitates redundant development efforts. To overcome this, we propose a standardized integration design paradigm and construct an open-source Python ecosystem built upon shared data contracts. By unifying long-format panel data and keyed forecast outputs while preserving model-specific implementations, the framework facilitates cross-engine hybrid modeling and rolling-origin evaluation. Furthermore, by integrating hierarchical weighting metrics with sparse reconciliation algorithms, we validate the multi-engine forecast reconciliation performance and memory efficiency on the M5 competition dataset. This approach effectively eliminates framework barriers and has achieved broad academic reuse and community adoption.

0 citationsRead paper

Recent publications

Latest Papers

Scale-Invariant Training for Time Series Foundation Models

Oct 05, 2026

This study addresses the issue of "scale contamination" in time series foundation model training, where scale inversion causes high-magnitude samples to dominate gradients. To overcome this, we propose a scale-invariant training method employing Reversible Instance Normalization (ReVIN). For homogeneous loss functions such as mean squared error, the loss is computed directly on scaled targets, ensuring scale-independent optimization trajectories. This work provides the first theoretical proof and correction of biases in existing normalization schemes, establishing a unified loss paradigm that requires only a single line of code for integration. Experiments demonstrate that our approach consistently reduces the MASE metric across four architectures, yielding average improvements of approximately 20% on GIFT-Eval and M-competitions, thereby significantly enhancing forecasting accuracy and robustness.

0 citationsRead paper

The Nixtlaverse: An Open-Source Ecosystem for Forecasting

Sep 30, 2026

This study addresses the incompatibility of data interfaces among statistical, machine learning, and neural network model families in forecasting software, which necessitates redundant development efforts. To overcome this, we propose a standardized integration design paradigm and construct an open-source Python ecosystem built upon shared data contracts. By unifying long-format panel data and keyed forecast outputs while preserving model-specific implementations, the framework facilitates cross-engine hybrid modeling and rolling-origin evaluation. Furthermore, by integrating hierarchical weighting metrics with sparse reconciliation algorithms, we validate the multi-engine forecast reconciliation performance and memory efficiency on the M5 competition dataset. This approach effectively eliminates framework barriers and has achieved broad academic reuse and community adoption.

0 citationsRead paper