This article explains how rule-based stock signals are produced and how to read one. It is educational and describes a research process, it is not personalised investment advice, and nothing here is a recommendation to buy or sell any security. Shishin publishes research; what an individual does with it is their own decision.
Search how do stock trading signals work and you get two incompatible answers wearing the same word. One is a person, or a black box, telling you to buy this, now. The other is a rule, applied to every name in a universe, on the same day, by the same arithmetic, that flags which ones meet a defined set of conditions. Both get called “signals.” They are not the same product, and the distance between them is most of what matters. This is how the second kind is actually built, and why a ranked, reproducible list is a fundamentally different object from a buy alert.
What a trading signal actually is
A trading signal, in the quantitative sense, is a flag: on a given date, a particular security meets a particular set of predefined, measurable conditions. That is the entire claim. Not “this will go up.” Not “buy this.” Only: the conditions the rule looks for are present on this name today, and here is the context around that fact.
The conditions are arithmetic. A moving-average relationship; a volatility band; a measure of how far and how fast a price has travelled; a liquidity floor so the name is actually tradeable. None of it requires an opinion, and none of it changes because the person running it is feeling bullish. Run the same rule over the same data tomorrow and you get the same answer. That reproducibility is the property that separates a signal from a hunch, and, as we will see, the property that most “signal services” quietly do not have.
So the honest one-sentence answer to how do stock trading signals work is: a rule reads the data, a name either meets the conditions or it does not, and the output is a record of which names did, ranked by how strongly. Everything else, the scoring, the classification, the regime context, is machinery for making that flag more informative. None of it is a promise about what happens next.
The “buy/sell alert” model, and why it is a different business
The dominant product sold under the word “signals” is the alert: a feed that pushes buy AAPL at 187, target 205, stop 181 and asks you to act. It feels like the more useful thing, it tells you exactly what to do. That is precisely the problem.
An alert collapses three separate jobs into one push notification: the research (is there a setup here?), the decision (do I want this exposure, at this size, in this account?), and the execution (when and how do I get in). The first is the only one a publisher can do honestly for a stranger. The second and third depend on your capital, your risk tolerance, your tax situation, and your other positions, none of which the alert sender knows. When a service performs all three for you, it is no longer publishing research; it is giving individualised instructions to people whose circumstances it has never seen.
The alert model also has a structural honesty problem. A pushed “buy now” is almost never accompanied by the full population of past calls, the ones that quietly didn’t work, or by a methodology you could reproduce. The winners get screenshotted; the losers get forgotten. There is rarely a point-in-time record, rarely a published drawdown, rarely a way to check whether the track record was assembled after the fact. It is a business built on the appearance of foresight rather than the evidence of an edge. The quant alternative is not better because it is fancier. It is better because every part of it can, in principle, be checked.
How a quant signal is actually produced
The pipeline is the same shape at every serious systematic shop, Shishin included. Five steps, each one mechanical, each one leaving a record.
- Scan. Start from a defined universe, a fixed, point-in-time list of eligible names, including the ones that later delisted or fell out, so the history is not quietly rewritten in the winners’ favour. Every name in that universe is examined the same way, every day. (Why the composition of that universe is the part most backtests cheat on is the subject of survivorship bias.)
- Enrich. For each name, compute the same set of indicators from its price and volume history, trend, volatility, range, thrust, distance from key averages, liquidity. This is the raw material the rest of the pipeline reads. It is arithmetic, not interpretation.
- Score. Collapse those indicators into a single comparable number. A good composite does not lean on any one input; it weights several so that no single indicator can carry a name on its own. That is the whole argument of composite scoring , eight features, one number, deterministic.
- Classify. A raw score is not enough, because the same number means different things in different contexts. A name that crossed its 50-day average yesterday is in a different state than one that has held above it for a year, and a bullish setup means something different in a broad advance than in a narrow, deteriorating tape. Two classifiers handle this: a per-name setup-state read, and a market-wide regime classifier that decides which posture is even appropriate today.
- Rank and publish. Sort the qualifying names by score and publish the list. Not “buy the top one” , just the ranked board, with the score and the context attached, so a reader can see why a name is where it is.
Shishin runs four scoring engines rather than one, because no single setup definition works across every market state. A breakout engine that prints money in a strong tape is the wrong tool in a defensive one; a different engine takes over there. A macro classifier decides which engine is allowed to fire on any given day, the architecture is described in four engines for four regimes. But the shape is the same as the five steps above. Scan, enrich, score, classify, rank. No step requires a forecast, and every step leaves a record you could audit.
Why a ranked list beats a buy command
The output being a ranked list rather than a single instruction is not a limitation. It is the point. A ranked board hands the reader the one thing an alert takes away: the decision. You can see the whole distribution, what scored highest today, how far the field drops off, which regime the system thinks it is in, and bring your own judgement, constraints, and risk appetite to it. The list is the research; the choice stays yours.
This is also why Shishin’s public feed deliberately does not carry entry prices, stop levels, or position sizes. Those are the three fields that would turn a research output into an instruction, and they are exactly the three that depend on the individual circumstances a publisher cannot see. The ranked score and the context are research anyone can use. A specific entry-stop-size triplet pushed to a stranger is advice dressed as information. We publish the first and withhold the second on purpose.
What a signal is not
Most disappointment with “signals” comes from expecting them to be things they are not. A signal is not a price target, the rule says a setup exists, not where the name will trade. It is not a guarantee; a high score is a statement about the present configuration, not a claim about the future. It is not an entry or exit command. And it is not personalised advice, it knows nothing about you.
Most signals, individually, also do very little. Across a long backtest the typical trade is neither a triumph nor a disaster; it lands in a quiet range either side of zero, and the edge comes from the shape of the whole distribution rather than any single call, the subject of the boring middle. Any service that implies each pick is a winner is describing a fantasy. The realistic claim is statistical: a small, repeatable tilt applied consistently across hundreds of names and hundreds of days. That is far less exciting than “buy this, retire early,” and it is the only version that survives contact with reality.
How to read a signal like a quant
If a signal is a flag plus context rather than an instruction, then reading one well means reading the context. Four things carry almost all of the information:
- The score, relative to the field. A composite of 72 means little in the abstract; it means a great deal if the rest of the board tops out at 60, and very little if half the list is above 80. Read it as a rank, not an absolute.
- The setup state. The same score on a name just emerging from a base is a different proposition than the same score on a name that is already extended. The state read is what tells you which.
- The regime. A strong-looking long setup in a deteriorating, narrow tape is worth less than the same setup in a broad advance. The regime is the backdrop the individual name is printing against.
- The indicator context. Volatility, distance from key averages, liquidity, the inputs behind the score. These tell you what kind of name it is, which shapes how much it could move in either direction.
Read that way, a signal is a well-organised starting point for your own work, not the end of it. Which is exactly what a research publication should be.
The honesty test: telling a real methodology from a marketing one
Because anyone can call anything a “signal,” the useful skill is not finding signals, it is telling the evidence-backed ones from the marketing. A handful of questions separate them, and each one maps to something a serious publisher will already have written down.
- Is there a published backtest, with the drawdown next to the return? A return quoted without the worst loss that produced it is uninterpretable. The two numbers, plus the risk-adjusted return, have to move together, the argument of the trinity, and the reason drawdown is treated as a feature to optimise rather than a side-effect to tolerate.
- Was the test run on a point-in-time universe?If the historical names were chosen with hindsight, only the companies that survived, the results are inflated before a single rule runs. See survivorship bias.
- Has the edge been significance-tested? A good backtest can still be luck. The bar quants actually use is far higher than the textbook one, precisely because so many strategies get tried before one looks good, the subject of statistical significance.
- Is anything verified forward, in the open? A backtest is a claim about the past. Watching the same rules run live, with every position and value visible, is the claim being tested in public, the case for paper-trading in public.
Notice that none of those questions is “what is your win rate” or “what was your best month.” Those are the numbers a marketer leads with. The questions above are the ones a methodology can only answer if it actually has one. A publisher who can answer all four is doing research. A service that can answer none is selling the appearance of it.
So: how do stock trading signals work?
A rule reads the data. Names meet the conditions or they do not. The ones that do are scored, placed in their per-name and market-wide context, ranked, and published. That is the whole mechanism, and its honesty comes from the fact that every step is reproducible and every step leaves a record. The output is not an instruction and not a forecast. It is a structured, ranked piece of research that points you at where to look, and then gets out of the way of the decision, which was always yours.
The alert model promises to make that decision for you. The quant model refuses to, on principle, because the decision depends on things the publisher cannot and should not pretend to know. That refusal is not a missing feature. It is the line between research and advice, and it is the line we publish on the correct side of.