dsh-bayes-predict

Most technical-analysis tools stop at qualitative calls—"MACD crossed above," "RSI is overbought"—without a way to fuse multiple indicators into a single probability answer. dsh-bayes-predict reframes stock prediction as a Bayesian inference problem: a prior built from the recent weighted win-rate is updated with empirical likelihood ratios drawn from MACD, RSI, moving-average alignment, Bollinger-band position, volume, momentum, and the 5-day trend, yielding a posterior probability of an N-day rally (with a 95 % confidence interval), a 0–100 score, and a three-state trend classification. Multi-horizon tables for 2/5/10/20 days let you compare confidence across time frames at a glance. Under the hood every signal is time-decay weighted (200-trading-day half-life), Laplace-smoothed with sample shrinkage, normalized by rolling volatility so thresholds adapt per ticker, and capped to a likelihood ratio of [¼, 4] to keep any single indicator from dominating. A layered prior mixes a market-wide likelihood table (market-prior.json) with per-stock statistics so small-sample tickers are not overwhelmed by noise. The data layer is entirely free and key-less: A-share daily bars come from Tencent's hfq endpoint (with Eastmoney as fallback), US bars from Tencent/Sina, and a built-in 30-minute TTL cache plus a hedge-adjustment break detector (single-day move > 8 % triggers an unadjusted cross-check) guard against silent data glitches. Zero third-party dependencies, Node 18+. It is aimed at quant enthusiasts who want a probability-driven single-stock view inside the DeepSeek Harness workspace, developers who want a clean, reproducible code base for studying Bayesian signal fusion in finance, and anyone who needs a sidebar one-click panel that returns a calibrated probability score without juggling API keys or heavy back-end stacks.