Five rounds on what open language models can and cannot do with financial text. In each round you first predict what will happen and lock your answer, then test it on the course data, and only then see the proof. Every answer comes from Qwen2.5 models run locally on the lecture data: Financial PhraseBank, finance tweets and the speeches of the European Central Bank (ECB).
Charts stay empty until you lock a prediction. In round 1, drag the ridge to turn it; scroll or pinch to zoom.
Round 1 · Temperature
- 1 Predict
- 2 Test
- 3 Prove
Round 2 · How big must the model be?
- 1 Predict
- 2 Test
- 3 Prove
Round 3 · Do examples help?
- 1 Predict
- 2 Test
- 3 Prove
Round 4 · Does the wording matter?
- 1 Predict
- 2 Test
- 3 Prove
Round 5 · Finding the right document
- 1 Predict
- 2 Test
- 3 Prove
Models: Qwen2.5-0.5B, 1.5B and 3B-Instruct (Qwen Team, 2024), run locally on an Apple-silicon laptop; no paid application programming interface. Financial PhraseBank (Malo et al., 2014), sentences with full agreement: 1,581 for training, 678 for testing; finance tweets: 900 test tweets. Retrieval: 400 ECB speeches (2022–2026) in 8,400 chunks of 200 words, dense embeddings bge-small-en-v1.5; question answering on the speeches from 2025 on. Code: lectures/L12/analysis_L12.py and build_interactive_L12.py on GitHub.

