How SignalFin uses AI

SignalFin's AI layer runs on Claude (Anthropic). This page explains exactly what each AI feature does, what data it is given, and the guardrails we apply. The design rule behind all of it: the AI reasons over numbers the platform computed — it never invents them.

What the AI does

1. Morning briefing. Each trading morning the AI writes a briefing with specific trade ideas — ticker, direction, entry, stop, and target — grounded in live pre-market gap data, relative volume, and the earnings calendar. The levels come from the platform's scanners; the AI's job is selection and reasoning, not making up prices.
2. Swing setup analysis. The setup scanner finds and stages candidates (entry zone, stop, target, risk:reward) from daily price structure. The AI writes the thesis for each one and flags what would invalidate it.
3. Tear sheet analysis. On any ticker, a written read of the fundamentals, technicals, sentiment, and news the tear sheet already displays — synthesized into a thesis you can check against the same numbers on the page.
4. Debate Mode. Three passes on one ticker: a bull case, a bear case, and a neutral referee that weighs both. Built to argue against you, not agree with you.
5. Watchlist analysis. On-demand research notes on the tickers you track.
6. Trade journal analysis. Reads your closed trades and reports patterns — which setups earn, which behaviors cost you, what to change. Statistics first, narrative second.
7. Portfolio commentary. Concentration and risk read-outs, upcoming-earnings context, and plain-language summaries of your connected holdings.

What the AI does not do

• It does not give personalized investment advice. Trade ideas are generated from market structure for a general audience. The AI does not know your finances, tax situation, or risk tolerance, and SignalFin is not a registered investment advisor.
• It does not place trades. SignalFin cannot execute, modify, or cancel orders on any connected account. Brokerage connections are read-only.
• It does not invent data. Every price level, statistic, and factor score the AI cites is computed by the platform from live market data and shown to you alongside the writing. If the data isn't there, the AI is instructed to say so rather than fill the gap.
• It does not guarantee outcomes. Setups lose. Stops exist because the thesis can be wrong. Any specific level is a structured plan, not a promise.

How it works

Each feature follows the same pipeline: the platform gathers structured data first (quotes, daily bars, factor scores, scanner output, sentiment reads, calendar events), then hands that package to Claude with a tightly scoped prompt. The model writes; the platform keeps the numbers. Outputs are cached — most research analyses for hours, briefings for the trading day — so two people asking about the same ticker at the same time see the same read.

Known limitations

• Language models vary between runs. Two analyses of the same ticker on the same day can differ in emphasis or conviction even with identical inputs. The structured levels stay comparable; the prose around them can shift.
• The AI only knows what the platform feeds it. Breaking news from the last few minutes, or data sources we don't integrate, are outside its view.
• Cached output can lag a fast-moving tape until the next refresh.

Privacy and data handling

Prompts contain market data and, for portfolio and journal features, the holdings or trades needed to do the job. They never include your name, email, or brokerage credentials. Data sent to Anthropic is processed under their API terms.