How SignalFin uses AI

SignalFin uses large language models to generate written commentary that accompanies the structured data in your portfolio view. This page explains exactly what the AI does, what it does not do, and the guardrails we apply to keep it useful and honest.

The short version: the AI describes what the structured analysis already shows, in plain language. It does not invent data, forecast prices, or recommend trades. Every claim it makes is grounded in the same numbers you can see for yourself.

What the AI does

The AI generates two kinds of output, and only these two:

1. Portfolio commentary. A short, structured read-out of your portfolio's key characteristics — concentration, recent movement, notable positions, upcoming catalysts. Always grounded in structured data shown alongside it.
2. Holdings explanations. When you drill into a position, a brief description of the company, its sector context, and the most recent material news. Sourced from public filings and news headlines.

What the AI does not do

This list is exhaustive, by design:

• The AI does not predict prices. No price targets, no “this stock will reach X.” If you ever see a number that looks like a forecast, it is reporting an analyst consensus from structured data, not a prediction.
• The AI does not give buy/sell recommendations. It describes what is in your portfolio. The decision to buy, sell, or hold belongs to you.
• The AI does not invent facts. Commentary is grounded in structured data we already have. It cannot reference earnings dates, revenue figures, or news headlines that are not in the underlying dataset.
• The AI does not have access to your personal information beyond your portfolio holdings. It does not know your tax situation, your goals, your risk tolerance, or your time horizon. Its output is structural commentary, not personal advice.
• The AI does not replace a financial advisor. Personal financial advice requires understanding your full situation. SignalFin is an analysis tool — useful, but limited by design.

How it works

Every AI-generated paragraph in SignalFin follows the same pipeline:

1. Structured data is computed first. Concentration scores, performance numbers, news headlines, earnings dates — all calculated deterministically from the underlying data.
2. The structured output is passed to the model. The model receives the actual numbers, names, dates, and headlines as context.
3. The model is prompted to summarize, not to add. System prompts explicitly instruct the model to describe only what is provided and to refuse to speculate, predict, or recommend.
4. Output is post-processed and validated. Numbers in the output are checked against the source data. Disclaimers are appended where appropriate.
5. Output is cached. Commentary is regenerated when the underlying data changes meaningfully — typically once per refresh cycle, not on every page view. This keeps cost and latency low and ensures consistency between sessions.

Guardrails

Specific rules the model operates under:

• No price predictions. The model is instructed to refuse forecasting. If the underlying data contains analyst consensus targets, those are reported with explicit attribution (“analyst consensus is X”) — never as SignalFin's view.
• No buy/sell language. Words like “buy,” “sell,” “you should,” or “recommend” are filtered. The model uses descriptive language only.
• No invented metrics. If a number is not in the source data, it cannot appear in the commentary.
• Citation of sources where relevant. When commentary references news, the underlying headlines and publications are linked alongside.
• Length-aware output. Analysis depth scales with portfolio complexity. A two-position portfolio gets short commentary; a fifty-position portfolio gets longer, more structured commentary. We do not pad short portfolios with filler.

Caching and consistency

AI output is cached for 24 hours per portfolio (matching the holdings refresh cadence). This has two purposes:

• Consistency. If you open the dashboard twice in one afternoon, the commentary is the same both times. This matters for trust — a different summary every time would feel arbitrary.
• Cost and latency. Generating fresh commentary for every page view would be expensive and slow. Caching keeps the platform fast and keeps the price predictable.

The cache is invalidated when holdings change meaningfully — a new position, a closed position, or a significant price move. Otherwise the commentary you saw this morning is the commentary you see this evening.

Privacy and data handling

• The model receives only the structured portfolio data needed to generate the commentary. It does not receive your name, email, or any identifier beyond an internal portfolio ID.
• Model providers are bound by data processing agreements that prohibit training on customer data. Your portfolio is not used to improve the underlying model.
• See the Privacy Policy for the full list of third parties that process data on our behalf.

Limitations and known issues

Be honest about what the AI is bad at, even within its narrow scope:

• Phrasing variation. Two portfolios with similar characteristics may get commentary phrased differently. The underlying assessment is the same; the language is not.
• Edge cases. Unusual portfolio compositions (e.g., single-position portfolios, all-cash portfolios) sometimes produce slightly awkward commentary. We're continuously refining prompts to handle these.
• News recency. The model only sees news that's in the source dataset. If a headline broke in the last few minutes, it may not be reflected in cached commentary.
• Ambiguous tickers. Some tickers map to multiple companies historically. We use current GICS classification, but edge cases exist.

Why we use AI here at all

A reasonable question. The structured data already exists — concentration scores, performance numbers, news headlines. Why generate prose around it?

Because data without context is hard to act on. A user looking at a 28% position concentration number wants to know: is that bad? for me? in this market? The commentary translates the structured signal into plain language, names the issue, and points to what changed since last time. That's a job humans are good at and software is bad at — except a well-prompted language model is, in this narrow case, also good at it.

The AI commentary is a feature, not a foundation. If you turned it off, every structured number, threshold, and chart in SignalFin would still be there. The AI simply describes what those numbers mean in language you'd use yourself.