deven367/temporal-stock-icl-2025

Temporal Stock Direction ICL 2025

CSV dataset for temporal direction experiments over 2025 monthly stock closes. It was built from Yahoo Finance chart data for 100 requested large-cap symbols; 98 symbols returned usable 12-month histories. The task is categorical: answer higher or lower for monthly close transitions.

Files

• —monthly_closes.csv: monthly close observations, 1,176 rows.
• —transitions.csv: adjacent monthly transitions, 1,078 rows.
• —train_chronological.csv, train_random.csv, train_reverse.csv: SFT rows with update + QA examples, 2,156 rows each.
• —train_chronological_qa.csv, train_random_qa.csv, train_reverse_qa.csv: QA-only SFT rows, 1,078 rows each.
• —val_historical.csv: direct historical direction questions, 1,078 rows.
• —val_latest.csv: latest-state direct questions, 490 rows.
• —val_icl_latest.csv: ICL prompts for only the final transition, 588 rows.
• —val_icl_all.csv: ICL prompts for every transition, 6,468 rows.
• —manifest.json: build metadata and label counts.

Columns

Common columns include prompt, completion, messages_json, symbol, date, previous_date, close, previous_close, category, task, and target. ICL rows additionally include context_order, prompt_id, and eval_scope.

Intended Use

Use this for quick EDA in pandas/HF Datasets, scaled ICL evaluation, and small SFTTrainer smoke tests. Chronological/random/reverse train files contain the same transition examples in different orders. EOF'