What is a black box stock rating?

It is a rating published as a number or grade with no visible breakdown of the inputs, weights or evidence behind it. You see the output but not the reasoning, which makes it impossible to verify, to date, or to reconcile with anything else you know about the company.

Is an explainable rating better than a black box one?

It is more usable, which is a different claim. Explainability tells you what a rating is made of and lets you check the documents yourself, so you can decide how much weight it deserves. It says nothing on its own about how well the rating is constructed.

What should a rating show me about insider buying?

It should name the filing, so you can see the transaction code, the date, the share count, the price and whether the Rule 10b5-1 box was ticked. Without that, a grant and an open market purchase look identical in a summary. The same applies to [unusual options activity](/learn/what-is-unusual-options-activity), where the contract details carry the meaning.

Why do ratings change without any news?

Usually because an input updated, not because anything happened at the company. New filings arrive, delayed volume data is published, an old event ages out of a lookback window, or the model is refitted. A breakdown showing what moved is the only way to tell which, a point covered further in [how AI scores stocks](/learn/how-ai-scores-stocks).

Learn

Why an unexplained rating is hard to use

Dan Seaton, FounderPublished 25 September 2026

A black box rating is a number with no visible reasoning attached. A stock gets an 84, or a four out of five, or a green badge, and there is nothing underneath it. You get the conclusion and are asked to take the working on trust.

The trouble is not that the model is complicated. Plenty of useful things are. The trouble is that a rating with no reasoning cannot be checked, argued with, or combined with anything else you already know. It is an instruction wearing the clothes of information.

An explainable rating does the opposite. It shows the components, names the evidence, dates it, and lets you look at the same material and reach a different view. That last part is the test.

A bare number leaves you with nothing to do

Say a stock is rated 84 today and was 61 last week. Without a breakdown you cannot tell whether the jump came from insider purchases, a burst of options activity, a price move that has already happened, or a quirk in how the rating handles thin volume. Each means something different, and some mean nothing at all.

You also cannot tell whether it is telling you something new. A rating that rises after a stock has already moved a long way is describing the move. One that rises because several officers filed purchases last week is describing an event you may not have seen. Same number, different use, no way to tell them apart from the outside.

Then there is scope. Every system weighs some things and ignores the rest, and a rating built on filings and flow has nothing to say about a pending court decision. If the scope is never published, you have no idea what silence means.

What explainability looks like in practice

Real explainability is not a paragraph of marketing about the model. It is four concrete things.

The first is a signal level breakdown. Rather than one figure, you see the components that make it up, which are active, and roughly how much each contributes. When the total changes you can see which part moved. That is the difference between "the rating went up" and "insider filings and off exchange volume both turned up, while price signals stayed flat".

The second is the source document. Every signal built on a filing should link back to it, and every signal built on market data should say what window it covers. If a system says insiders have been buying, you should be able to open the Form 4 and read who, when, how many shares, at what price.

The third is honest timing. Public data carries lags that vary by source, and an explainable rating shows both dates: when the event happened and when it became public. The SEC requires a Form 4 before the end of the second business day following the transaction. Congressional reporting is slower, with a periodic transaction report due by the earlier of 30 days from the filer being made aware of the trade or 45 days from the trade, which is why tracking politician trades means reading delayed information. 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.

The fourth is a stated scope. A short, plain account of what the rating uses and what it ignores is worth more than any description of the algorithm.

The filing tells you what the label cannot

Form 4 records each transaction with a code, and the codes are not interchangeable. The SEC defines P as a purchase of securities on an exchange or from another person and S as a sale. A is a grant, award or other acquisition from the company, M is the exercise or conversion of a derivative received from the company, F is payment of an exercise price or tax liability using part of the securities received, and G is a gift. A rating that treats every acquisition as buying reads routine equity compensation as enthusiasm, and one that treats every disposal as selling reads tax withholding as an exit. Only the code separates them, so only a system that shows you the filing lets you check. Our explainer on what insider buying means goes further into this.

Form 4 also carries a checkbox for transactions made under a contract, instruction or written plan intended to satisfy the affirmative defence conditions of Rule 10b5-1. A sale made under a plan set up months earlier reads differently from a discretionary one, and the box is right there on the form.

Congressional disclosures have their own texture. The reporting threshold is transactions over $1,000, and amounts are given in brackets starting at $1,001 to $15,000. A rating that converts those ranges into a single number is making a choice for you, and unless it shows the range you cannot see what that choice was.

Explainable does not mean simple

A model can be complex and still produce an audit trail, and a simple one can still be opaque if nobody will show it to you. Explainability is a property of what gets published, not of how many parameters sit behind it.

It is also worth being sceptical about contribution figures. When a system reports that a signal contributed a given share of a rating, that describes the model's internal arithmetic, not the world. It tells you what the system weighted, which is useful, but it is not evidence the signal caused anything. The filings and market data underneath are the evidence. The attribution is a map of the model.

Questions worth asking of any rating

What inputs does it use, and what does it ignore?
What is the timestamp on each input, and how delayed is the slowest?
Is a number relative to the stock's own history or to the whole universe?
What would make it change tomorrow, and can I see what changed yesterday?
Can I open the underlying filing or data behind any claim it makes?

A system that answers all five is showing its work. One that answers none is asking for trust it has not earned. Most sit somewhere between, and knowing where is the point of asking.

How InsiderPulse handles this

InsiderPulse scores assets from 0 to 100, and the score is meant to be opened rather than accepted. Contributing signals are shown separately, drawn from SEC Form 4 filings, US congressional disclosures, options activity, dark pool and volume data, prices and news, so you can see which part of the evidence is active. Ask Pulsey answers questions about a score with citations, so a claim can be traced back to its source. The score summarises what the data currently shows, it is not a recommendation, and it does not cover ASX shares.

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