minicpm5-stock-v2-forward-return

# MiniCPM5 Stock v2 — Forward-Return Labels Binary BUY/SELL stock-direction dataset where labels come from **actual forward 5-day returns** (BUY > +2%, SELL < -2%, middle band dropped), not news sentiment. All features are strictly causal (no look-ahead): last 20 daily returns, RSI(14), volume ratio vs 20d MA, 20d volatility, 5d/20d momentum, 20d relative strength vs SPY. - `train_minicpm5_v2.jsonl` — 5,056 rows, 16 tickers, class-balanced - `val_minicpm5_v2.jsonl` — 1,586 rows, 4 held-out tickers (TSLA/NFLX/AMD/WMT), class-balanced - `minicpm5_v2_forward_return.ipynb` — ready-to-run Colab/Kaggle T4 training notebook (LoRA on MiniCPM5-1B, <1 hr, includes confidence-filtered eval) ## Honest information ceiling (measured, 5 seeds, dense stride-1 sampling) GradientBoosting baseline on the identical features, held-out tickers: | Coverage (confidence-filtered) | Accuracy | |---|---| | 100% of signals | 50.9% | | top 50% | 52.2% | | top 30% | 54.2% | | top 20% | 56.2% +/- 1.5 | Expect LLM results in this neighborhood. Full-coverage accuracy near 60% on this feature set is not achievable — configurations that appeared to reach 57-58% in sweeps were sampling artifacts (sparse-stride phase alignment) and did not survive dense re-evaluation. The honest headline metric is accuracy at reduced coverage (the model abstains on low-confidence samples), which is also how real trading desks operate. Each row: `prompt` (text), `label` (BUY/SELL), `ticker`, `date`, `fwd_return`, and `features` (raw numerics, so classical baselines can run on identical rows).