DISC Learns to Combine Time-Series Similarity Metrics and Predict Transfer Success

Researchers have introduced DISC (Domain-Integrated Similarity Cost), a learnable framework that calibrates and combines similarity metrics for time-series datasets. Its goal is to predict in advance how well a forecasting model will transfer from source data to target data. The main result: the integrated score agrees better with actual forecast errors than any single metric used as a baseline on the reported split.

Why estimate transfer in advance

Time-series forecasting is a workhorse in energy, retail, finance and infrastructure monitoring. When a team has several candidate sources for pretraining or fine-tuning, the question is which data to train on to get the best results on the target. Only brute force gives an honest answer: adapt the model to every source–target pair and compare the errors. That is expensive, especially when there are many candidates.

Hence the need for a quick estimate. Many dataset similarity metrics have been proposed for this purpose, but they measure different things and correlate differently with what actually matters: forecast quality after transfer. The DISC authors measured this gap directly. They compared source–target similarity scores with downstream forecast errors across different setups, including transfer learning.

No single metric works for every task

The analysis showed that no single metric is a universal indicator. How useful similarity is depends on the task. A metric that reflects transfer well in one setup may correlate only weakly with error in another. Choosing a metric by hand carries a high risk of getting it wrong from the start.

DISC addresses this by learning from observed transfer data. The framework:

• selects which similarity metrics are needed for a given task at all;
• assigns them weights;
• combines the metric values into a single score;
• calibrates the final score against actual forecast quality.

The key idea is complementarity. Different metrics capture different aspects of the relationship between datasets, and combining them more accurately separates the differences between source and target that affect model behavior after transfer from those that do not.

What comes out

After calibration, DISC scores source–target pairs using only the similarity metric values. No forecasting models need to be fine-tuned for this. According to the authors, on the reported split this score agrees more closely with downstream results than the individual baseline metrics do.

The second finding concerns the field itself: time-series similarity is a task-dependent concept. A learned integration of metrics is a stronger proxy for transfer behavior than hand-picking a single metric.

Where this is useful

For teams that work on forecasting and reuse data across domains, DISC offers a concrete workflow. First, compute a set of similarity metrics for the candidates. Then get the integrated score and shortlist the few most promising sources for actual adaptation. A short check replaces the full brute-force search.

There is a limitation: calibration relies on observations of transfer, so the quality of the estimate depends on how closely new tasks resemble those the framework was tuned on. The results are reported for one split, and it is worth checking how well they hold up in your own domains. For practitioners assembling corpora for pretraining time-series models, the methodological framing is valuable in itself: measure similarity not by a familiar metric, but by how well it predicts the final error.

The paper is available on arXiv under number 2610.31288.