TRANSPARENT BY DESIGN
How TradeMates turns market data into an analysis.
See the sources, model passes and safeguards behind a report — and the exact places where data can be unavailable or a review can time out.
Evidence first. Interpretation second. Checks before delivery.
TradeMates does not ask one model for an unstructured opinion. It freezes a structured market-data snapshot, sends that same input through three focused production passes, applies deterministic safeguards, then runs a separate second-AI quality review.
What happens in each stage
Collect and normalize evidence
The system identifies the listing, currency and instrument type. It loads available quotes, history, filings, TTM metrics, estimates, earnings, analyst revisions, insider activity, options, news and peers from the configured providers. Fields are normalized while the input is built; missing fields remain unavailable.
Run three focused production passes
The frozen stockData payload is sent to three structured model calls in parallel: core thesis and risk, business fundamentals and valuation, and technical/market context. Each pass interprets the same snapshot, rather than fetching a different view of the market.
Apply safeguards and review
The system normalizes scenario probabilities, clamps invalid ranges, replaces model-reported confidence with a deterministic value and reconciles the verdict with the score. A separate second AI model then checks the assembled report before it is shown.
The named data sources
The exact fields available depend on the listing, instrument type, market and provider response. TradeMates keeps unavailable evidence visible instead of silently filling gaps.
Financial Modeling Prep Premium
Company profile, historical statements and TTM data, valuation and DCF cross-checks, estimates, analyst-grade revisions and earnings-call transcript context where available.
Finnhub
Live quote and market context, company and earnings signals, analyst and insider data where the configured endpoint provides coverage.
Yahoo Finance
Supplementary quote, price-history and market-reference data used when available for the requested instrument and field.
Provider responses are merged into the input by field. The current report does not claim a provider badge for every individual field because that provenance is not stored field by field.
What the score actually means
Investment Score is a conservative 0–100 estimate of the probability that the investment gains value within 90 days. It is not a general company-quality rating, a target price or a guarantee. Scores above 80 are intentionally rare.
The displayed verdict is rule-based
After the production model returns its score and recommendation, a deterministic reconciliation applies the canonical thresholds. If the model recommendation disagrees with the score, the displayed verdict is corrected to the matching rule.
Production score vs. shadow experiment
In production, the model returns the score and its own factor breakdown. The factor names and weights in that breakdown are model output; there is no fixed seven-factor production weighting guarantee. The fixed seven-factor weighting belongs to shadow-v0.1.0, a deterministic experiment used for comparison and not used to replace the production score.
Confidence is calculated after the model pass
The displayed confidence value replaces the model’s self-reported confidence. It is calculated from data completeness (25), signal agreement (30), financial history depth (15), analyst coverage (15) and earnings predictability (15), with small premium bonuses for transcript coverage and five or more analyst-grade revisions.
A second AI review stays visible
A separate lightweight Gemini-family model cross-reads the finished analysis against the frozen data. It checks verdict consistency, fair-value plausibility, thesis balance, evidence thinness, data contradictions, scenario consistency and catalyst grounding.
Reviewed · clean
The second model returned no quality flags.
Reviewed · flagged
Minor or major issues are shown as badges; the report is still delivered. Flagged does not mean withheld.
Review unavailable
A timeout or skipped review is labeled explicitly. The main analysis is delivered unchanged; unavailable is not presented as clean.
An inspectable report trace
This shows the kind of chain a user can inspect in a report. Values here are illustrative and carry no live provider timestamp or field-level provenance claim.
Where scheduled work runs
Interactive reports use the TradeMates analysis pipeline. Scheduled non-email prewarm jobs run to completion on the Hetzner worker first; only the finished result and its metadata are then sent to Supabase. Email prewarms use their separate path. A completed analysis can be reused for 60 minutes, and a cache hit is accepted only when the current stock-data fingerprint matches.
The time horizons are intentionally different
What TradeMates does not do
- It does not guarantee a future return or remove investment risk.
- It does not turn missing data into a confident number. ETFs, ETPs, funds and other instruments can have different coverage or calculations.
- It does not present a historical backtest as a promise. Backtesting is a separate frozen-snapshot validation layer.
- It does not replace your own decision, suitability check or professional advice.
Now inspect the process on a stock.
Enter a ticker and follow the evidence, score, safeguards and quality signals in the finished report.