PRICE CONTEXT
60 saved observations¥45.45Last saved close · 3 Nov 2026
Historical close (CNY)Current forward-adjusted series
FROM NEWS TO A TESTABLE FORECAST
Luxshare Precision立讯精密 · SZ 002475
Follow the evidence. Inspect the model. Understand the limits.
i
A historical replay, ready to inspect. The four candidate events were first recorded after the final price observation. They do not affect this forecast. Opening this page does not fetch new prices or call an AI model.
THE RESEARCH TRAIL
One case, four steps.
- 6 source excerptsHuman-selected news with URLs and timestamps
- 4 candidate eventsExtracted in a saved OpenRouter run; review pending
- 0 active event inputsAll four arrive after the price-data cutoff
- 130 validation observationsSame dates and horizon for each baseline
Explore the demo
02 / SOURCES & EVENTS
Every claim has a trail.
Six excerpts organized into five evidence groups. Two buyback reports describe one underlying event.
Review boundary. Kezia’s notes check source articles against manual annotations. They explicitly do not accept or reject the four exported Agent events. Candidate-event approval remains pending.
03 / MODEL EVALUATION
Compare before you conclude.
| Method | MAE ↓percentage points | RMSE ↓percentage points | Direction matchnegative / flat / positive | 90% interval coverage | Mean interval widthpercentage points |
|---|
Lower error is better. “—” means the baseline does not produce an interval. The zero-return baseline always predicts flat; its direction score is not bearish accuracy. No method here is a trading strategy.
Inspect the five largest model errors
| Decision date | Target date | Predicted return | Actual return | Absolute error |
|---|
04 / SCOPE & LIMITATIONS
What this version can tell you.
Timing is part of the evidence.
Prices end on 3 November. The source excerpts were first recorded after 19:37 on 5 November (China time). Those events cannot be inserted into an earlier decision date.
Event coefficient: 0.0000. Forecasts match the price-only comparison. News-driven improvement has not been established.
An interval is not a guarantee.
The adaptive interval uses rolling standardized-residual quantiles; it is not simply a fixed Gaussian band. Validation windows overlap and are not independent trades.
The model has no execution simulation, fees or daily price-limit handling. Results do not establish profitability.
A saved research case.
One supported stock; 443 price observations; six selected excerpts. Original source claims have not been independently verified, and news coverage is not exhaustive.
No live refresh or open-ended chat is included. The Agent run used NVIDIA Nemotron via OpenRouter; this page makes no AI calls.
Review tools · demonstrate application states
These are labeled UI simulations for review, not live service tests. They never change the saved research results.