Launching today
Rivonai-an InvestGPT
AI for China stocks: a research engine, not a GPT wrapper.
3 followers
AI for China stocks: a research engine, not a GPT wrapper.
3 followers
InvestGPT turns global headlines into China stock signals — backed by real research. It pulls from a growing knowledge base of capital market research (equity reports, supply chains, fundamentals) indexed in a vector database (RAG). When news breaks, it returns a specific signal: which stocks, direction, severity, time horizon, reasoning chain. Then you talk to it — drill into any thesis, ask follow-ups. The knowledge base grows daily. Free daily credits. Premium unlocks unlimited use.
Love that you can drill into the reasoning chain, that's what separates this from a typical alert bot. One thing I'd love to see is a backtest view for each signal showing how similar past setups played out, so I can gauge how the model has historically handled comparable headlines before sizing a position.
@cafermqyf
Thank you for the thoughtful feedback — and you've zeroed in on exactly the right distinction. The reasoning chain isn't a nice-to-have; it's the difference between trusting a signal and just reacting to one. You should understand why something fired, not just that it fired.
Your backtest view idea is excellent, and honestly it's something we've been actively exploring. Here's how we're thinking about it:
1. Historical Signal Matching
For each new signal, we'd surface the N most comparable past setups — matched on headline type, market regime, and signal characteristics — then show how those actually played out across multiple time horizons (1h / 4h / 1D). You'd see the distribution of outcomes, not just a single anecdote.
2. Model Performance Attribution
Not just "did the trade work," but how the model's reasoning held up. Did the factors it weighted heavily turn out to matter? Were there blind spots it missed? This turns each signal into a learning loop, not just a bet.
3. Confidence Calibration
Over time, this builds something you can actually trust: when the model flags "high conviction," how often is it right? That track record maps directly to position sizing — which is really the point you're making. A signal without historical context is just an opinion; a signal with calibrated history is actionable intelligence.
This shifts each alert from a binary go/no-go into a data-informed decision.
One question that would help us prioritize the design: would you find this more useful as a per-signal drill-in panel (context right when you're reviewing a specific alert) or as a standalone analytics dashboard (broader model evaluation across many signals)? Both have merit — just curious which fits your decision workflow better.