"AI analysis" is one of the most abused phrases in finance. Often it means a large language model was asked to summarise a press release. That's not what we do. This piece walks through the eight quantitative modules that actually produce a FinsightAI verdict, in plain English.
The architecture
Each analysis runs a stock through eight independent modules. Each module produces a score from 0 to 100, where 50 is neutral. The final composite is a weighted blend, with weights that shift slightly by asset class (a small-cap stock leans more on quality and value; a large-cap with active options leans on sentiment and options flow).
1 — Trend
Measures the structural direction of the price series. We use a blend of the 50-day and 200-day moving averages, the slope of each, and the position of price relative to both. A stock above a rising 200-day with a rising 50-day above it scores in the 70s or 80s. A stock below a falling 200-day with the 50-day cutting through scores in the teens.
2 — Momentum
Measures the rate of change, not the direction. We use a smoothed 14-period RSI, weekly return distribution, and MACD histogram dynamics. Momentum and trend disagree often — a stock can be in a long uptrend but with weakening momentum (typically a yellow flag), or be in a downtrend with stabilising momentum (sometimes the first sign of a base forming).
3 — Volatility
Measures risk character. We compute realised volatility over multiple windows (10, 30, 90 days) and compare to the stock's own history. Importantly, this module does not penalise high volatility per se — it flags changes. A normally-quiet utility suddenly trading with the volatility of a small-cap miner is more informative than a small-cap miner trading like itself.
4 — Value
For equities, the value module looks at P/E versus the stock's own 5-year median and versus its sector. P/B for financials. EV/EBITDA where relevant. The output is not "is this cheap in absolute terms" but "is this cheap relative to its own history and its peers." Absolute-value calls require macro context we don't pretend to have.
5 — Quality
Returns on equity, debt-to-equity, gross margin stability, and free cash flow conversion. Quality scores tend to be sticky — they don't move much month-to-month — which is part of their value. A high-quality business going through a temporary price drawdown is a very different setup from a low-quality business doing the same thing, and the model treats them differently.
6 — Sentiment
News sentiment from major financial wires, scored over rolling windows. We use a fine-tuned classifier rather than a generic LLM because finance has its own vocabulary (a "miss" is bad, a "beat" is good, "guidance" is loaded). This module is the noisiest of the eight; we down-weight it for thinly-covered stocks where news flow is sparse.
7 — Options
For stocks with liquid options chains, we read implied volatility, put-call ratios, and skew. Options traders pay for information; their positioning sometimes leads the spot market. For stocks without meaningful options (most PSX names, smaller crypto), this module is skipped and the others reweight.
8 — Macro
The wrapper. For a Pakistani bank, the relevant macro variables are the SBP policy rate, inflation, and rupee stability. For a US tech name, they're 10-year yields, the dollar index, and credit spreads. For Bitcoin, they're real yields, ETF flows, and Fed liquidity. Each asset gets a macro context score that adjusts the composite up or down.
Putting it together
The eight module scores feed into a final composite. We then map the composite to one of seven verdict bands, from "Strong Sell" to "Strong Buy," with confidence levels based on how much the modules agree. When modules disagree sharply (e.g. strong trend but weak quality and stretched valuation), the verdict comes back as "Mixed" with a flag — and that's often the most useful output of all.
What this is not
It's not a crystal ball. It's a structured way to look at a stock in ten seconds that would otherwise take a careful analyst an hour. It will be wrong sometimes — every model is. The goal isn't to be right on every call; it's to give you a consistent, transparent process so your decisions get better over time, not worse.