What actually happens when a model scores a stock

A stock score is a number standing in for a pile of evidence. Behind it sits a set of measurements about a company and its stock, each one converted into a comparable value, then combined by weights into a single figure on a fixed range such as 0 to 100.

That is most of the answer. A model does not look at a ticker and form a view about the business. It reads the features it has been handed, compares each reading against that stock's own history and against other stocks, applies weights and returns a summary.

So the questions that matter are mechanical ones. What is being measured, how much each measurement counts, what the model was fitted on, and what the resulting number is actually claiming.

Features are the raw measurements

A feature is one measurable thing expressed as a number. The count of insider purchases made on the open market in the last 30 days is a feature. So is the size of a purchase set against what that person already held, the ratio of call volume to put volume, the share of a session's volume printed away from the lit exchanges, or the gap between today's price and its average over the past 50 sessions.

Raw numbers are rarely comparable across stocks, so they get rescaled before they are any use. A purchase worth a few hundred thousand dollars is routine at one company and remarkable at another, and ten million shares traded is quiet for one ticker and extraordinary for the next. Most systems convert each raw reading into a rank or a percentile, either against the stock's own past or against the wider universe on the same day. What the model then sees is not "eight insiders bought" but "this is unusually high for this company".

Every feature also carries a window. Thirty days of insider filings is a different feature from ninety days of the same filings, and the two can point opposite ways at once. Windows are a quiet design choice, because changing one changes the score without anything changing at the company.

Weights decide what counts

Once features are comparable, something has to decide how much each contributes. There are two broad approaches and most real systems mix them.

Weights can be set by hand, which is analyst judgement written down. Someone decides a cluster of open market purchases by several officers counts for more than a single purchase by one director, and the formula reflects that. The reasoning is legible, but it is still somebody's opinion.

Weights can also be learned, by fitting them against historical data and a chosen target. The target is the part people skip over, and it defines the whole exercise. Fitting to "did this stock rise over the following month" produces a different model from fitting to "did it have an unusually large move in either direction". Two systems can use identical inputs, learn different weights and disagree completely, purely because they were pointed at different questions.

Training data is the whole of the model's experience

A model knows nothing outside the data it was fitted on. That bounds it in three ways: the period covered, the universe of assets covered, and the honesty of the timestamps.

Timestamps are where most of the damage happens. Everything in training has to be dated by when it became knowable, not when it happened. The SEC requires a Form 4 before the end of the second business day following the transaction, so the trade and the public record of it sit on different dates, and a model using the transaction date is training on information nobody could have acted on. Congressional disclosures widen the gap further, since a periodic transaction report is due by the earlier of 30 days from the filer being made aware of the trade or 45 days from the trade itself. FINRA publishes weekly off exchange volume on a two week delay for Tier 1 NMS stocks and a four week delay for Tier 2 NMS stocks and OTC equity securities, so a feature built on it must be lagged the same way in training as in use.

Get those dates wrong and the model learns from a version of history that never existed, which is invisible from the outside unless somebody publishes the methodology. Our note on how the InsiderPulse score works sets out which inputs feed the number and how they are timed.

A score summarises evidence, it does not forecast

This distinction does most of the work. A score is a compression of the present, not a statement about the future. The honest reading of a high number is "a lot of the things this system tracks are active at once right now", and the honest reading of a low one is that they are not. Neither is a claim about what happens next.

That framing also explains why the number alone is thin. Two stocks can carry the same score for unrelated reasons, one built on a run of insider filings and the other on options and volume behaviour. Same figure, different evidence. Which is why the component breakdown matters more than the headline, a point covered in black box versus explainable ratings and in the overview of AI stock pickers.

The limits, stated plainly

Regime change is the big one. A model fitted across a particular stretch of market conditions absorbs that stretch's habits, and when conditions shift the relationships it encoded can weaken or invert. Nothing in the score announces this. The number keeps arriving with the same confident formatting.

Data quality is the mundane one. Filings get amended, tickers change, forms arrive late and corporate actions distort price histories unless they are carefully adjusted. Congressional disclosures report amounts in brackets such as $1,001 to $15,000 rather than exact figures, so any feature built on trade size works with a range rather than a value.

Survivorship bias is the quiet one. If the historical universe a model was fitted on contains only companies that still exist today, every failure has been deleted from its experience and it has learned about a friendlier world than the real one. That happens whenever a data set is assembled looking backwards instead of recorded forwards.

None of this makes scoring useless. It makes a score a starting point for reading rather than a conclusion, which is also the short answer to whether AI can predict stocks.

How InsiderPulse handles this

InsiderPulse scores every asset it covers from 0 to 100 using SEC Form 4 insider filings, US congressional trade disclosures, options activity, dark pool and volume data, prices and news. The score is a summary of what those inputs currently show, not a forecast and not a recommendation. Each score can be opened up to see which signals are contributing, and Ask Pulsey answers questions about one with citations back to the source. Coverage is US listed stocks and ETFs plus crypto, forex and commodities, so ASX shares are not included.

InsiderPulse is a data tool. Nothing on this page is financial advice.