TiRex — Zero-Shot Forecasting
- Pretrained time-series model
- No target-dataset training or fine-tuning
- Univariate closing-price history
- Context length up to 2,048 observations
- Nine forecast quantiles from 0.1 to 0.9
Bachelor Thesis · 2026
Zero-Shot vs. Supervised Learning in Financial Time-Series Forecasting
Researcher & Machine Learning Developer
Can a pretrained forecasting model generalize to a new financial domain without fine-tuning? I evaluated TiRex, a zero-shot time-series model, against supervised baselines trained specifically for the target data with engineered technical features.
10German large-cap stocks
2dataset variants
3forecast horizons
5walk-forward folds
4compared models
Academic context: This project is a forecasting comparison, not a trading system, a claim of reliable future-price prediction, or financial advice.
01 — Research Question
Can zero-shot forecasting compete?
Can TiRex match or outperform established baselines even though these models are trained specifically on the target dataset and use multivariate input features?
The comparison is intentionally asymmetric. TiRex receives only the historical closing-price series and is used without fine-tuning. LightGBM and ElasticNet are trained on the target data and receive engineered multivariate features.
02 — Comparison Setup
Less target-specific information,
the same evaluation.
Supervised Baselines
- Trained separately for each horizon
- Multivariate engineered features
- Returns and lag features
- SMA 5, SMA 10, SMA 30 and RSI 14
- Relative distance to moving averages
- LightGBM and ElasticNet
03 — What I Built
End-to-end research pipeline.
- Collected and prepared daily stock-price data
- Built the feature-engineering pipeline for supervised models
- Integrated and validated the TiRex zero-shot pipeline
- Implemented LightGBM and ElasticNet baselines
- Implemented expanding-window walk-forward evaluation
- Prevented look-ahead leakage with chronological splits and horizon gaps
- Evaluated point and probabilistic forecasts
- Performed paired Wilcoxon tests with Holm correction
- Created result tables and visualizations
- Documented and interpreted the experimental results
04 — Data & Evaluation
Chronological by design.
Daily price data from ten German large-cap stocks across multiple industries. The long dataset covers January 2015 to June 2026; the shorter dataset covers June 2023 to June 2026.
Forecast targets
Models
Primary metrics
Additional metrics
Statistical method
Before applying TiRex to stock data, the local evaluation pipeline was checked against selected published GiftEval benchmark results.
05 — Evaluation Design
Expanding-window walk-forward evaluation.
The models were evaluated using five chronological folds. Every test period occurred strictly after its training data to prevent information leakage.
Training counts can vary by model because feature preparation changes the number of usable observations; tooltips show the source-data range.
06 — Results
Overall Model Comparison
Values below 1 indicate lower error than the naive baseline in the descriptive aggregated comparison.
07 — By Horizon
Performance by Forecast Horizon
One chart, two dataset variants, three horizons. Toggle individual models below.
* For normalized WQL, ElasticNet is descriptive only — no genuine quantile forecast.
08 — Stock Level
Stock-Specific Probabilistic Performance of TiRex
Mean fold-level normalized WQL, grouped by stock, dataset, and horizon.
Values below 1 indicate better probabilistic forecasting performance than the naive baseline. All displayed averages are below 1; the strength varies across stocks and horizons.
09 — Statistical Evidence
Matched fold-level comparisons.
Negative mean differences favor TiRex; positive differences favor the comparison model. Color is supplemented by position, labels, and tooltip text.
10 — Key Findings
What the evidence supports.
Strong Against Trained Baselines
TiRex achieved lower point-forecast errors than LightGBM and ElasticNet in the main statistical comparisons.
Best Result in Uncertainty Modelling
TiRex achieved an overall normalized WQL of 0.803 and significantly outperformed LightGBM with quantile regression in both dataset variants.
Robust Across Most Horizons
Fourteen of eighteen horizon-specific comparisons significantly favored TiRex. The only significant baseline advantage was LightGBM at the one-day horizon in the long dataset.
The Naive Baseline Remained Strong
Aggregated normalized MASE was 0.950, but fold-level comparisons showed that the naive baseline remained difficult to beat consistently for point forecasts.
11 — Interpretation
Competitive with less target-specific information.
The results show that a pretrained forecasting model can compete with supervised models despite using less target-specific information. TiRex’s clearest advantage was probabilistic forecasting and uncertainty modelling. For point forecasts, however, the naive baseline remained a strong benchmark.
12 — Scope and Limitations
Boundaries of the study.
- Ten selected German large-cap stocks
- May not generalize to international markets, small caps, or other asset classes
- Two historical dataset variants
- 1-, 10-, and 30-trading-day horizons
- No trading strategy was evaluated
- No transaction costs, slippage, or risk management
- Forecast accuracy does not imply profitability
- No causal claims about financial markets
Disclaimer: This project presents an academic forecasting comparison and does not constitute financial advice.