A 0.6B model, six baselines, one controlled evaluation

Context and role

The problem

High-volatility markets are hard to forecast at multiple horizons at once. A model can look good at one step and produce incoherent paths over 24 hours.

Most reported improvements are measured against weak or inconsistently tuned baselines. A genuine comparison needs a controlled protocol and the same evaluation for every model.

A small backbone is cheap to run, but only useful if it can represent multi-timeframe structure and produce coherent trajectories rather than independent point predictions.

Approach

• Built a multi-timeframe input pipeline ingesting OHLC data across 1-hour to 1-day windows.
• Fully fine-tuned a Qwen3-0.6B Transformer backbone, rather than training a small task head on frozen features.
• Designed a trajectory-aware hybrid loss combining step-wise and compounded-path objectives, so the 24-hour return trajectory stays coherent.
• Evaluated under the controlled FinTSBridge protocol against six established deep-forecasting baselines.
• Compared on multi-step information coefficient (msIC), not a single cherry-picked horizon.

What I did not claim

The boundary of a project is part of its evidence. These are the claims this work does not support.

• This is thesis research, not a client deployment or a trading system.
• No claim of profitability or live trading performance.
• Not a general-purpose chatbot or instruction-following model.
• The result is a measured improvement under one protocol, not a universal claim about small models.

Results

• The fine-tuned Qwen3-0.6B achieved the highest multi-step information coefficient against six established deep-forecasting baselines.
• The predicted price paths correlated with realised prices under the FinTSBridge evaluation.
• The work produced a reusable pattern: freeze the evaluation, train the small model, and compare against tuned baselines under identical conditions.
• The project is the evidence behind the fine-tuning and evaluation services on this site.

Stack

• Python
• PyTorch
• Hugging Face Transformers
• Qwen3-0.6B
• FinTSBridge

What this means for you

The useful part of this project is not the artifact. It is the method: define the measurement first, keep the scope honest, and hand over something your team can run. That method is what transfers to a new problem.

If your situation looks similar, the related services below are the place to start. If it does not, send a short description and I will tell you whether I am the right person.