AI-Powered Stock Advisory: What It Actually Does, and What It Cannot
“AI-powered” has become the default adjective in Indian financial marketing, usually attached to a moving-average crossover. Here is where machine learning genuinely helps in market research, where it fails predictably, and the four questions that expose the difference.
What “AI-powered” usually means in practice
The phrase has no regulatory definition and no technical floor. In Indian retail financial marketing it is currently applied to at least five very different things:
| What is claimed | What it often is | Actually useful? |
|---|---|---|
| “AI-powered signals” | Rule-based indicator crossovers with a modern label | Sometimes — but the label adds nothing |
| “Machine learning models” | A classifier fitted to historical price data | Only if validated out-of-sample |
| “Deep learning predictions” | A neural network trained on a small, noisy dataset | Usually overfitted |
| “AI risk management” | Volatility-scaled position sizing | Genuinely useful — and decades old |
| “Sentiment AI” | Keyword scoring of news or social feeds | Occasionally, as one input among many |
None of this means quantitative methods do not work. It means the label is not the evidence. A provider should be judged on how its models are tested, not on which fashionable noun appears in its marketing.
Where machine learning genuinely helps
There are parts of the research process where statistical modelling has a real and durable advantage over human judgement:
- Consistency of execution. The largest measurable edge in most systematic strategies is not signal quality — it is that the rules get applied identically every time. Models do not move stops because a position “feels” like it will come back.
- Scanning breadth. Evaluating hundreds of instruments against a defined condition set, every session, without fatigue or attention bias.
- Regime classification. Identifying whether current conditions resemble those in which a strategy historically performed — a genuinely hard statistical problem where models beat intuition.
- Volatility-adaptive sizing. Continuously adjusting position size to keep rupee risk constant as volatility changes. Simple, mechanical, and consistently valuable.
- Non-linear interaction detection. Finding conditions where several weak signals combine into something meaningful — patterns that are difficult to specify by hand.
- Honest measurement. Computing expectancy, drawdown, cost drag and per-regime performance across a full history, without the selective memory humans apply to their own records.
Notice that most of these are about discipline and measurement, not prediction. That is where the real value sits, and it is the least marketable part.
Where it fails, predictably
- Low signal-to-noise. Financial time series are overwhelmingly noise. Models trained on noise learn noise, and do it very convincingly.
- Non-stationarity. The market's statistical properties change. A model trained on one volatility regime can be actively harmful in another, and it will not tell you that it has stopped working.
- Small effective sample sizes. Ten years of daily data is about 2,500 observations — a small dataset by machine-learning standards, and one in which the observations are not independent.
- Reflexivity. A widely-discovered edge is arbitraged away. Markets are not a fixed system being measured; they respond to being traded.
- Silent failure. A model that has stopped working produces confident outputs indistinguishable from a model that is working, until the P&L reveals it.
- Data problems. Survivorship bias in historical constituent lists, corporate-action adjustment errors, look-ahead leakage from data timestamped incorrectly. These produce spectacular backtests and catastrophic live results.
The overfitting problem, in plain terms
Given enough parameters, any model can be tuned to describe any historical dataset almost perfectly. This is not skill; it is curve-fitting. The tuned model has memorised the past rather than learned anything generalisable, and it will fail as soon as it meets data it has not seen.
The defences are well established and easy to ask about:
- Out-of-sample testing. Parameters set on one period, results measured on a period the model never saw during tuning.
- Walk-forward analysis. Repeated rolling optimise-then-test cycles, so the reported result is a chain of genuinely out-of-sample outcomes.
- Parameter stability. A robust strategy performs reasonably across a range of parameter values. If it only works at exactly 14 periods and collapses at 13 or 15, it has found an accident.
- Realistic cost and slippage assumptions applied inside the test, not subtracted afterwards.
- Forward, live performance — timestamped, unedited, including losing months. This is the only evidence that cannot be manufactured after the fact.
Four questions that expose the difference
- 01“Is this live performance or a backtest?”If backtest — was it walk-forward, and over what period? A single in-sample curve is not evidence.
- 02“What are the model's inputs, in plain language?”Exact parameters can reasonably be proprietary. The category of input cannot. “It's AI” is not an answer.
- 03“In which conditions does it lose money?”Every model has hostile conditions. A provider who cannot name theirs has not looked, or will not say.
- 04“What is the maximum drawdown, and when?”With dates. A model presented as having no bad period has had its bad periods removed from the presentation.
What regulation says about automated research
Using models does not create a separate regulatory category. A provider publishing research generated by quantitative methods is still a Research Analyst under the SEBI (Research Analysts) Regulations, 2014 if it publishes research, or an Investment Adviser under the 2013 regulations if it gives personalised advice. The same obligations apply:
- No assured or guaranteed returns — the technology used is irrelevant to this prohibition.
- No custody of client funds or securities.
- The advertisement code applies, including its prohibition on superlative claims and misleading use of past performance. “Our AI delivered 300% returns” is precisely the kind of claim it addresses.
- Disclosure obligations attach to the research, whether a human or a model produced it.
Separately, requirements around algorithmic order routing, API-based trading and retail algo access have been tightened in recent years. If a provider proposes to place orders in your account automatically, that is a different activity from publishing research and carries its own regulatory considerations — verify the current framework before agreeing to anything of the sort.
“Our AI has an 85% accuracy rate.” Accuracy without average win and average loss is uninformative — see the expectancy problem in our intraday guide. A system can be 85% accurate and lose money consistently.
How EqtPulse uses quantitative methods
EqtPulse is a SEBI Registered Research Analyst (Reg. No. INH000028565). Our research is quantitative and rules-based. We would rather describe what that means than lean on the label:
- Signals are generated by defined conditions — measurable, repeatable, and applied identically every session rather than by discretionary conviction.
- Risk is calibrated to volatility, so position size contracts as the instrument's range expands and rupee risk stays within a defined band.
- Strategies are separated by regime dependence — momentum, mean-reversion, option-selling, delta-neutral and commodity desks behave differently in the same market, which is the point of running more than one.
- Performance is published forward and complete — 1, 3, 6 and 12-month windows with maximum drawdown, profit factor, average win, average loss, trade count and full monthly P&L including the negative months.
- Every strategy is labelled high risk, because every one of them is.
We do not claim our models predict the market. They estimate conditions under which a defined edge has historically been present, and they apply risk control consistently. That is a smaller claim than most of this industry makes, and it is the one we can support.
Frequently asked questions
What is AI-powered stock advisory?
In principle: research where statistical or machine-learning models are used to identify patterns, generate signals, size positions or manage risk, rather than relying solely on human discretion.
In practice, the label is applied to a very wide range of things, from genuine quantitative research infrastructure to a spreadsheet with two indicators. The label itself carries no information — the testing methodology behind it does.
Can AI predict the stock market?
No model predicts market prices reliably. What well-built statistical models can do is estimate conditional probabilities — that under a defined set of conditions, one outcome has historically been somewhat more likely than another — and act on that edge consistently across many trades.
The distinction matters. A system built on 'we predict the market' collapses at the first regime change. A system built on 'we exploit a small statistical edge with strict risk control' can survive one.
Is AI-based trading allowed in India?
Systematic and algorithmic trading operate within the existing regulatory framework, with requirements that apply to exchange-level algorithmic order routing and to registered intermediaries publishing research. Using quantitative models to generate research does not exempt anyone from the Research Analyst or Investment Adviser regulations, the advertisement code, or the prohibition on assured returns.
Requirements for retail algorithmic trading and API-based order routing have been tightened in recent years — confirm the current position from the latest SEBI and exchange circulars.
Why do AI trading systems fail?
Most commonly, overfitting: the model was tuned until it described historical data beautifully, and that description does not generalise. Other recurring causes are look-ahead bias in the data, unrealistic execution assumptions, ignoring transaction costs, and regime change — the market behaving in ways not represented in the training period.
Almost none of these failures are visible in a backtest. They appear in live trading, which is why forward performance is the only evidence worth much.
Is an AI advisory better than a human analyst?
Different failure modes, not better or worse. Models do not get bored, frightened or vengeful, and they apply rules identically at 9:20am and 3:20pm. They also cannot recognise that the current situation has no historical precedent, and they fail silently rather than expressing doubt.
The strongest setups usually combine systematic execution with human oversight of regime and risk — not one replacing the other.
Can an AI advisory guarantee returns?
No. The prohibition on assured returns applies regardless of the technology used to generate research. Any provider offering guaranteed profits — with or without an AI label — is making a claim that no registered intermediary is permitted to make.