Tsqa Eval

---
name: tsqa-eval
description: Evaluates large language models on time series question answering across five tasks: forecasting, imputation, anomaly detection, classification, and open-ended reasoning. It probes the model's ability to integrate textual context with numerical time series data for both precise numerical prediction and natural language explanation. Use when the user wants to benchmark on TSQA, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2503.01875
  bibtex_key: kong2025timemqa
  confidence: high
---

# tsqa-eval

> Time-MQA: Time Series Multi-Task Question Answering with Context Enhancement — Kong et al. (2025) (arXiv:2503.01875, 2025)

## What this evaluates

Evaluates large language models on time series question answering across five tasks: forecasting, imputation, anomaly detection, classification, and open-ended reasoning. It probes the model's ability to integrate textual context with numerical time series data for both precise numerical prediction and natural language explanation.

## Datasets

- **TSQA** — total 200000; splits: test (250)

## Metrics

- `average MSE` — range: [0, ∞)
  - Mean Squared Error averaged across all forecasting and imputation instances. Calculated as the mean of squared differences between predicted and actual values. Lower values indicate better performance.
- `accuracy` **(primary)** — range: [0, 1]
  - Percentage of correctly predicted labels or answers across anomaly detection, classification, judgment (true-false), and multiple-choice questions. Calculated as correct predictions divided by total instances. Higher values indicate better performance.

## Input / output format

**Input**: Time series data paired with a natural language question/query. Questions and answers are clearly labeled and tokenized.

**Output**: Numerical predictions for forecasting/imputation tasks; class labels for anomaly detection/classification; natural language answers with reasoning for open-ended reasoning (judgment and MCQ) tasks.

## Scoring recipe

```python
def compute_metrics(predictions, gold, tasks):
    results = {}
    for pred, gold_val, task in zip(predictions, gold, tasks):
        if task in ['forecasting', 'imputation']:
            results.setdefault('average MSE', []).append((pred - gold_val) ** 2)
        else:
            results.setdefault('accuracy', []).append(1.0 if pred == gold_val else 0.0)
    return {k: sum(v)/len(v) for k, v in results.items()}
```

## Common pitfalls

- The test set is extremely small (only 50 samples per task), making reported metrics highly susceptible to sampling variance.
- Forecasting tasks use long time series, which the authors explicitly note leads to relatively high MSE values compared to imputation tasks.
- Open-ended reasoning combines multiple-choice and true-false formats but evaluates both solely by accuracy, masking potential differences in question difficulty.

## Evidence (verbatim from paper)

> For evaluation, we randomly selected 50 QA pairs for each task type (or question format). Forecasting and imputation tasks were evaluated using average MSE, while anomaly detection, classification, and open-ended reasoning tasks (including multiple-choice questions (MCQs) and true-false questions (Judgment)) were measured by accuracy. A lower value of MSE ↓ and a higher value of accuracy ↑ indicate better performance.

## Citation

```bibtex
@misc{kong2025timemqa,
  title={Time-MQA: Time Series Multi-Task Question Answering with Context Enhancement},
  author={Kong et al. (2025)},
  year={2025},
  note={arXiv:2503.01875}
}
```

- arXiv: 2503.01875