Guide & Free Tool · Machine learning
AI in trading: what machine learning can and cannot do in Indian markets
The short answer
Markets are close to the hardest setting anyone has ever pointed a statistical learner at, and the reasons are structural rather than a matter of finding a better model. The signal is thin, the process producing it keeps changing, and overfitting is the default outcome of any search rather than an error you avoid by being careful. Where machine learning genuinely earns its keep for a retail trader is not prediction at all. It is the unglamorous middle: classifying conditions, ranking and screening, checking execution, auditing your own record, and reading text. So the useful question is never can AI predict the market. It is which small, checkable job am I giving it.
This page refuses the usual pitch. It contains no forecasts, no claimed accuracy or return figures for any system, and no product names, because every one of those is a way of dodging the question of whether the underlying problem is tractable at all. What it contains instead is arithmetic. Every figure below is computed from a stated simulation whose assumptions are written out, and several of them produce numbers that are uncomfortable for the industry that sells this technology to retail traders. The tool immediately below does the same job for a specific claim: give it the headline numbers of any strategy and it will tell you how much of them a plain search over that many variations would have produced with no edge at all.
Free interactive tool
Backtest Reality Check
Enter the headline claims of any AI or ML strategy: its backtest Sharpe, how much data it was fit on, how many variations were tried, how many trades it made, what the model was asked to do, and how it was validated. The tool estimates how much of that Sharpe pure search luck would produce, flags the classic red flags, and reads it as a plausible edge, unproven, or marketing fluff. It judges the claim, not the future.
Start from a scenario
The claim
The annualised Sharpe ratio the backtest reports.
How long a history it was tested on.
Parameter sets, features and thresholds tested before keeping the best.
Roughly how many separate bets the record contains.
0%LUCK-
EXPLAINED
EXPLAINED
Why the number can lie
Your Sharpe against the search-luck band
The bar is the Sharpe scale from zero. The shaded band on the left is the Sharpe the best of your variations would show with no real edge at all. The white marker is your claimed Sharpe; the green gap, if any, is the edge that sits above pure luck.
Your Sharpe Explained by search luck Edge above luck
Red flags in this claim
This tool checks one claim against one idea, the deflated Sharpe, which is only the first of four gates. A backtest is a hypothesis, not a track record; proving an edge, rather than merely finding one, is the discipline that the method we teach is built around, and the whole of Stage 4, the Quantitative Edge, is spent on it.
Start here: a beautiful equity curve is the default output of a search
Before any argument about machine learning, there is a fact about searching that most people never internalise, and it explains more failed strategies than every other cause combined. If you try many variations of a rule and keep the one that performed best, the winner's record is not a measurement of skill. It is the maximum of a distribution. And the maximum of enough noisy draws is always impressive, whether or not anything real is present.
The figure below is that statement made concrete. It simulates 200 variants of a trading rule, and every single one of them was built to have a true edge of exactly zero. There is no signal anywhere in the data. The variants are then split at the two-year mark: the first two years stand in for the history you search over, the second two for the money you eventually trade. The best variant over the search window gained 58.7 index points and posted an annualised Sharpe ratio of +1.68, a record most retail traders would consider strong evidence of something.
Notice what the last number does. The expected best Sharpe from a search is not a mystery to be discovered after the fact. It can be calculated in advance from two inputs you already know: how many variations you tried and how long a history you tried them on. That is the idea behind the deflated Sharpe ratio, and it is precisely what the tool at the top of this page computes. The comparison that matters for any strategy claim is not the headline number against zero. It is the headline number against what your own search would have produced from nothing.
Here is where machine learning enters, and why it makes this worse rather than better. A person running a spreadsheet search knows roughly how many variations they tried. A model does not present you with that number. Every additional input, every hyperparameter setting, every architecture you tried and abandoned, and every quiet re-run after a disappointing result is another trial, and almost nobody writes any of them down. The effective search breadth behind a modern model is typically orders of magnitude larger than the number its builder would report, which means the luck threshold it has to clear is far higher than the one they are checking against.
A backtest is not evidence until you know how many backtests you ran to get it. The number of attempts is part of the result, and discarding it is the single most common way an honest person produces a dishonest number.
This failure has a large literature and a well-mapped anatomy, and it is not the only way a historical test flatters a weak idea. Look-ahead leakage, survivorship in the universe, unadjusted corporate actions and unmodelled costs each do their own damage, and most retail backtests carry several at once. The full catalogue, with the mechanism and the countermeasure for each, is set out in the companion guide on backtesting integrity and the eight ways a backtest lies, and this page does not repeat it. What follows instead is why markets in particular are so hostile to the method, which is a different question and a deeper one.
What a model is, and what markets refuse to supply
Strip away the vocabulary and a machine-learning model is a function fitter. You give it inputs and a target, and it searches for the shape that connects them in the data it can see. That is an extraordinarily powerful idea, and in the domains where it has transformed everything, four conditions happened to hold at once. There was far more signal than noise. The relationship being learned did not change while you learned it. Independent examples were abundant and cheap. And you could check an answer immediately, for nothing.
Markets supply none of the four generously, and that is the whole of the argument on this page. It is worth laying the four side by side, because the contrast is what makes the difficulty structural rather than incidental. No amount of model sophistication changes a single row of the table below, which is why the phrase a better model would solve this is almost always misplaced.
| What the method needs | A well-behaved domain | A market | Consequence for you |
|---|---|---|---|
| Signal that dominates noise | A picture of a cat is almost entirely signal. The label is essentially determined by the pixels. | Day-to-day price changes are dominated by noise. The forecastable component of a return is a thin tilt, not a determination. | Evidence requirements are enormous, and a short record carries almost no information. |
| A relationship that holds still | The rules connecting pixels to objects do not change while you study them, and nobody adapts to your discovery. | Participants adapt, market structure and costs change, and volatility regimes turn over. A relationship can weaken, vanish or invert. | A model fitted on the past is fitted to a world that has partly ended, so it decays instead of compounding. |
| Abundant independent examples | Millions of labelled images, and more can be collected on demand. | A decade of daily bars is roughly 2,500 observations, and overlapping windows are not independent, so the effective count is smaller still. | Model capacity has to stay very small, which is the opposite of how these tools are usually sold. |
| A cheap, immediate check | The answer is verifiable at once, at no cost, as often as you like. | You learn whether a forecast was right only after the market moves, one observation at a time, and one observation proves nothing either way. | You cannot iterate quickly without the iterations themselves becoming the search that fools you. |
Read the right-hand column downward and you have the four sections that follow. The next three take the first three rows in turn and put a number on each, because the abstract version of this argument is easy to nod along to and easy to forget the moment a convincing-looking curve appears on a screen. The fourth row, the absence of a cheap check, is the one that quietly determines which jobs are worth giving a model at all, and it is the organising idea of the section on where machine learning does earn its keep.
The signal is thin, and thin signals demand enormous evidence
The first structural problem is the ratio of signal to noise. Decades of published work on return predictability converge on an uncomfortable summary: the part of a return that is forecastable from information available beforehand is small, and predictors that look convincing within a sample frequently fail outside it. That is not a fringe position or a counsel of despair. It is the mainstream finding, and it is what makes the honest version of this subject so different from the marketed one.
It also explains why this page publishes no accuracy figures, and why you should treat any that are advertised to you as uninformative rather than impressive. There are three separate reasons. On a target that is close to a coin flip, an accuracy number is dominated by estimation noise, so the difference between an ordinary result and a spectacular one is often just sample size. Accuracy also ignores the sizes of the moves you get right and wrong, so a system can be correct most of the time and still lose money steadily, and the reverse is equally possible. And it is trivially inflated after the fact by choosing the period, the universe or the decision threshold once you have seen the outcomes. A figure that can be manufactured three different ways is not a measurement.
The useful way to think about a thin edge is not accuracy at all. It is how long you would have to run it before its record could distinguish it from nothing, and that has an exact answer. For an annualised Sharpe ratio, the t-statistic of a track record is approximately the Sharpe multiplied by the square root of the number of years. Turn that around and the years of data needed to reach a given standard of evidence scale as the square of the ratio between that standard and the true Sharpe. The figure below draws both halves of the problem.
The left panel is the part that costs people money. Every one of those fourteen paths is the same strategy, with the same genuine advantage, differing only in the order the good and bad periods arrived. Yet after five years the luckiest and unluckiest lives are separated by roughly 159 percentage points, and two of them are still under water. A trader living any one of those lives has no way, from the record alone, to tell which one they are in. This is why abandoning a rule after a bad quarter and adopting one after a good quarter are both, statistically, coin flips dressed up as judgement.
The right panel is the part that should reset expectations about what a backtest can prove. The two-year test that flattered the champion in the previous section is not merely weak evidence. For any edge a retail trader is likely to have, it is close to no evidence at all. And the higher of the two thresholds is not an invention: when a research literature has collectively tested hundreds of candidate factors, the conventional t-statistic of two stops being an appropriate bar, because with that many attempts a fair number of false results will clear it by chance. Work on multiple testing in the cross-section of expected returns has argued that a newly reported result should clear roughly three instead. If you have run a broad search, you are in the second column, not the first.
None of this means an edge cannot exist or cannot be measured. It means the measurement is a long, deliberate exercise in expectancy and sample size rather than something a backtest settles in an afternoon, and the mechanics of doing it properly, in R-multiples and net of costs, are set out in the guide on how to measure a trading edge. What the arithmetic here establishes is the prior you should walk in with: a strong claim built on a short record is far more likely to be a search artefact than a discovery.
The ground moves, which is why models in markets decay instead of compounding
The second structural problem is that markets are not stationary. In a well-behaved domain, a model compounds: more data makes it better, and better is permanent, because the thing being learned does not change. A market offers the opposite bargain. The relationship you fitted is a property of a particular set of participants, rules, costs and volatility conditions, and all four of those turn over. Participants adapt to whatever is working. Market structure changes: settlement cycles, lot sizes, surveillance measures, index composition and the cost stack have all shifted in India within the span of an ordinary backtest. And any edge that becomes widely known is traded until it is not worth trading.
The practical consequence is unintuitive and worth stating plainly. In markets, a model is a perishable good. Its expected performance is highest on the day it is fitted and declines from there, and the decline is not a sign that something went wrong. It is the normal behaviour of a fitted relationship in a system that keeps moving.
The bottom strip deserves a second look, because it undercuts the standard remedy. The usual answer to non-stationarity is to detect the change and refit. But the rolling estimate crosses zero 14 times over a decade in which the true relationship reverses only 2. Most of those crossings are the estimator's own noise, not the world turning. A trader who refits every time the measurement flips sign will spend the decade chasing an artefact of their own estimation window, arriving at each new regime slightly after it has ended. Detecting a regime change in real time is a genuinely hard problem, and it is hard for the same reason everything else here is hard: the signal is thin relative to the noise.
There is a design lesson buried in this, and it is one of the few places where the honest answer is also the practical one. If the relationship you are fitting is unstable, you should want a model that is coarse, slow to change, and built on a mechanism you can state in a sentence. A rule with two parameters that expresses something you understand about how the market works will survive a regime turn better than a flexible model with forty inputs, not because simplicity is a virtue in itself, but because there is less of it to be wrong. It also fails visibly rather than quietly, which is worth more than it sounds.
What this rules out. It rules out the compounding story that makes AI attractive elsewhere, the one where the model keeps learning and keeps improving. In markets the honest planning assumption is the opposite: whatever you fit will decay, you will not know precisely when the decay began, and the budget for re-validation is a permanent operating cost rather than a one-off setup task. Any offer built on the premise that a model gets better on its own over time is describing a domain that is not this one.
More model, more mirage: why power is a liability here
The third structural problem is overfitting, and it needs separating from the search problem in the first section, because they are different mechanisms that arrive at the same wreck. The first section was about breadth: trying many rules and keeping the best. This one is about capacity: how much freedom a single model has to bend itself around the data it was given. Both manufacture a result out of nothing, and a model with a lot of capacity does the second one automatically, without you having to try anything at all.
The figure below is a plain ordinary least squares fit, deliberately, because the effect has nothing to do with any exotic technique. A model is given one year of daily observations and up to 24 inputs. Only the first two carry any real signal, and even they are weak: together they explain 3.1 per cent of the variance. Every other input is pure noise, constructed to be unrelated to the target. The whole exercise is averaged over 200 independent draws, and the honest score is measured on 20,000 observations the model never touched, so neither curve is the luck of a single run.
Two things follow, and the second is the one that inverts the usual sales pitch. First, the amount of model you can afford is set by how much data you have, not by how much modelling you feel like doing. On one year of daily bars the answer is roughly three inputs. Three. Anyone describing a system with dozens of features fitted on a couple of years of history is describing the right-hand edge of that chart, whether or not they know it.
Second, a more powerful model does not escape this. It arrives at the mirage sooner. A gradient-boosted ensemble or a neural network has vastly more effective capacity than 24 linear coefficients, which means it can fit the noise in the same 250 observations far more thoroughly and produce a far more beautiful in-sample curve. In a domain rich in signal and examples, that capacity is exactly what you want. Here it is a liability, and the standard intuition that a stronger tool should give a better result is precisely backwards.
The regularisation trap. The correct textbook response is to constrain the model: penalise complexity, select features, tune with cross-validation. All of that genuinely helps, and all of it is itself a search. If you choose the penalty strength, the feature set or the validation scheme by looking at the result on your held-out data, that data has become training data and your honest estimate is gone. The tuning has to happen strictly inside the fold, and the number of tuning attempts has to be counted along with everything else, or you have simply moved the overfitting one level up where it is harder to see.
Where machine learning genuinely earns its keep
Everything so far has been about why one particular job is close to impossible. That job is prediction: asking a model to tell you what a price will do. It is worth being clear that this is not a verdict on the technology, because there is a set of jobs in a retail trading process where a model does real work, and they have almost nothing in common with forecasting.
What separates them is not cleverness and not the amount of data alone. It is three properties considered together. Are there examples to learn from? Can you check whether an answer was right without waiting on the market? And what does a wrong answer cost, money or merely time? Prediction fails all three at once. The jobs below pass on at least two, and the ones that pass on all three are where a retail trader should start.
Take those in turn, because the differences matter more than the list. Execution quality is the best-behaved problem in the set. Every order you place generates a labelled example within seconds: what you intended, what you got, and the difference. The target is mechanical rather than strategic, the examples accumulate quickly, and a wrong answer costs you a slightly imperfect order rather than a position. A model that learns when your orders fill badly is doing ordinary, well-posed statistics on data you own.
Regime classification is the most useful and the most easily abused. Useful, because knowing whether conditions are trending or choppy changes which of your rules should be active and how large you should be, and getting that roughly right is worth more than getting a forecast precisely right. Abused, because the data is thin, so the moment a regime model has a dozen states and twenty inputs it has become a return predictor wearing a different label. Two or three states, a handful of inputs, and a stated mechanism.
Screening and ranking is the least glamorous and the most reliably valuable. You are not asking what will happen. You are asking a machine to apply criteria you wrote down, consistently, across a universe too large to inspect by hand, and to do it the same way on a Friday afternoon as on a Monday morning. The edge is consistency, not insight, and consistency is genuinely hard for a human and trivial for a machine.
Anomaly detection in your own trading record is the one almost nobody does and the one with the most immediate payoff. Your journal is a labelled dataset about a single, highly non-random agent: you. A model that flags trades which do not match your written rules, or that surfaces the conditions under which you break them, is learning something stable, checkable and entirely within your control. Nothing about it depends on the market being predictable.
Text processing is where large language models are legitimately strong, with one condition attached. Pulling named fields out of a filing, grouping announcements by type, or turning a long document into a structured summary are real tasks with abundant examples and immediate verification. The condition is that you check the output against the source. A model that reads a document for you and is right most of the time is useful; a model you have stopped checking is a rumour with a confident tone.
| The job | What it replaces | Why it is learnable | How you check it | Where it stops |
|---|---|---|---|---|
| Execution quality | Eyeballing whether a fill felt bad | Every order is a labelled example, and the target is mechanical | Compare intended against achieved on the next order | It cannot tell you whether the trade was worth taking |
| Regime classification | An unstated feeling about conditions | The target is observable and low-dimensional, and a wrong call is cheap | Does its state label match what you would have said, on old data | Add states and inputs and it silently becomes a return predictor |
| Screening and ranking | Inconsistent manual filtering | You supply the criteria, so nothing has to be discovered | Re-run the same criteria by hand on a sample | It ranks against your rules; it does not know if your rules work |
| Anomaly detection in your record | Not noticing your own drift | The subject is one consistent agent, and the data is yours | Read the flagged trades and judge them yourself | It finds deviations from your process, not flaws in it |
| Text into structured fields | Reading everything, or reading nothing | Language is the one domain here with abundant examples | Open the source document and verify the field | Extraction is not interpretation, and it is never a view |
One rule cuts across all five. Whatever the model does, it goes through a forward test before it touches size, because the failure modes on this page are invisible in historical data by construction. A simulated run on data that arrives after the model was built is the only test that cannot be contaminated by hindsight, and what a simulator can and cannot prove, including the several ways it flatters you, is covered in the guide on paper trading and what it actually proves.
A language model is not a market model
The arrival of fluent general-purpose assistants has scrambled this conversation, so it is worth being precise about what they are. A large language model is trained to continue text plausibly. That is the objective it was optimised for and the thing it is genuinely extraordinary at. It has no internal model of an order book, no representation of a price as a quantity, and no mechanism connecting anything it has read to what a market will do next.
Asking such a system what a price will do is therefore not a hard question badly answered. It is a category error. And the reason it is dangerous rather than merely useless is that the model will answer, fluently, with structure and caveats and an air of consideration, because producing text that reads like careful analysis is exactly the skill it has. The confidence of the output carries no information about the reliability of the content. Fluency is not evidence.
The specific hazard is code, not prose. Ask an assistant to write a backtest and you will usually get something that runs, looks professional and contains a subtle timing error: a signal computed from the same bar it trades on, an indicator that quietly uses the full series to compute a rolling value, a label that encodes information from after the decision point. Every one of those produces a beautiful equity curve and a live failure, and none of them looks wrong on the screen. Fluent code with a look-ahead bug is more dangerous than no code, because it comes with the authority of having worked. Read every line that touches the time index, and test the logic on data where you already know the answer.
Used within its actual competence, the same tool is genuinely valuable, and it would be silly to pretend otherwise. Explaining a concept you have half-understood. Reviewing code you wrote and asking what could be wrong with it. Turning a vague intuition into a written specification precise enough to test, which is often the single most useful thing you can do with a trading idea. Summarising a long document into fields you then open the document to verify. Drafting the rules of a process so that you have something concrete to argue with. In every one of those the model is working with language, the answer is checkable, and being wrong costs a few minutes.
The line is clean enough to state as a test. If the question has a right answer you can verify today, an assistant is a reasonable place to start. If the question's answer only arrives when the market moves, you have left the domain the tool was built for, and no amount of prompting brings you back into it.
The rulebook, as it stood on 18 July 2026
The regulatory picture matters less than most coverage suggests, but two points are worth knowing, and both are easy to get wrong because the widely repeated versions are inaccurate.
The binding provision is not a new AI framework. Chapter IIIB of the SEBI (Intermediaries) Regulations, inserted by an amendment notified on 6 February 2025, places responsibility squarely on the regulated entity: an intermediary using artificial-intelligence or machine-learning tools, whether it built them or bought them from a third party, is solely responsible for the privacy, security and integrity of the data, for the output produced, and for compliance with applicable law. The practical effect is that outsourcing the model does not outsource the accountability, which is the correct principle and one worth applying to your own use even though it does not bind an individual trader.
The broader guidance is still a proposal. SEBI published a consultation paper on the responsible usage of artificial intelligence and machine learning in the securities markets on 20 June 2025, setting out five core guiding principles: model governance, investor protection and disclosure, a testing framework, fairness and bias, and data privacy and cyber-security measures. As of 18 July 2026 it had not been finalised into a circular or a regulation. If you have read elsewhere that it establishes six principles named ethics, accountability, transparency, auditability, data privacy and fairness, that list does not appear in the paper; it is a garbling that has spread through secondary commentary. Read the source.
Where the algo rules live, and when to re-check. The retail algorithmic trading framework, including the identification requirements, broker responsibility and the routing obligation on algo providers, is a separate subject with its own guide on algorithmic trading in India and the four things that phrase means, and it is not repeated here; it became applicable to all stock brokers on 1 April 2026. Separately, the Digital Personal Data Protection Act, 2023 was operationalised by rules notified in November 2025, but the substantive obligations on data fiduciaries, including notice, consent and breach reporting, commence on a phased timetable and were not yet in force on 18 July 2026. Every date and requirement on this page should be confirmed at the SEBI and MeitY websites before you rely on it, because this area is moving quickly.
The question worth asking
Put the four structural problems together and a rather freeing conclusion falls out. The question that dominates the marketing, can AI predict the market, is not a hard question with an unknown answer. It is a malformed question, because it asks about the one job that fails every condition the method requires. Replace it and the whole subject becomes tractable: which small, checkable job am I giving this thing, and how will I know if it is doing it badly?
Five questions settle that, and they are worth writing down before any code is written, because each one is far harder to answer honestly after you have seen a result you like.
| Ask | Why it decides everything | The answer that should stop you |
|---|---|---|
| 1. What exactly is the label, and at what instant does it become knowable? | Nearly every leak is a label that quietly contains information from after the decision point. Naming the instant is what exposes it. | Any answer that needs a sentence beginning with "well, roughly" |
| 2. How many genuinely independent examples does a year produce? | It sets the ceiling on how much model you can afford, at roughly one input per hundred observations before the gap eats the signal. | Fewer than a few dozen, unless you have decades of history |
| 3. Can I check an answer without waiting for the market? | A checkable job can be iterated safely. An unverifiable one turns every iteration into another draw in a search you are not counting. | No, I would have to wait and see |
| 4. What does a wrong answer cost, and is it recoverable? | It sets how much autonomy the model gets. Cheap reversible errors can be automated; expensive irreversible ones need a human in the loop. | A position, immediately, with no natural check before it is placed |
| 5. How many things did I try before this one, and did I write the number down? | Without the count, the headline result cannot be compared against what search luck alone would produce, so it means nothing. | I did not keep count |
Question five is the one that gets skipped, and it is the one the tool at the top of this page exists to enforce. Counting your attempts feels bureaucratic in the moment and it is the difference between a number that means something and a number that means nothing. Keep a plain text file. Write down each variation before you run it. The discipline costs nothing and it is the only thing standing between you and the first figure on this page.
Finally, notice what the five questions have in common. Not one of them is about the model. They are all about the problem you are pointing it at, which is where the difficulty actually lives. Someone who answers these five carefully and then fits something deliberately simple will do better than someone who skips them and reaches for the most powerful tool available, and that ordering is not a moral preference. It is what the arithmetic in the four figures above says.
Where this page stops, honestly
Three limits, stated plainly, because a page arguing that other people overclaim has an obligation not to.
The figures here are simulations, not measurements of the Indian market. Each one is generated from a process whose assumptions are written into the caption, and the numbers they produce are outputs of those assumptions. What they demonstrate is that certain effects follow necessarily from the structure of the problem: that the maximum of many noisy trials looks impressive, that a thin edge needs a long record, that a fitted relationship can invert, that added capacity buys apparent fit and costs real fit. Those conclusions do not depend on the particular numbers, and you can verify each one with a spreadsheet and an afternoon.
Second, some institutions do extract real value from machine learning in markets. They do it with data a retail trader cannot buy, infrastructure a retail trader cannot build, research teams whose full-time job is validation, and portfolios of many weak signals combined at a scale where thin edges become viable. That is a genuine existence proof for the technique. It is not evidence that the same technique transfers to one person, one laptop and a few years of daily bars, and the marketed version of this subject relies almost entirely on blurring that distinction.
Third, and most importantly: automation does not manufacture an edge, it executes whatever process you already have, faster and more consistently. A SEBI study published in September 2024 found that about 93% of individual traders in equity derivatives made net losses over FY22 to FY24, with aggregate net losses exceeding ₹1.8 lakh crore over the three years, and a follow-up study in July 2025 found the proportion broadly unchanged. No technology has moved that number, because the constraint was never processing speed. Adding a model to a process without an edge produces the same outcome with better logs.
Educational content, not advice. This page explains what machine learning can and cannot do in a market setting. It is not a recommendation to trade, to invest, or to use any particular technique, and it is not investment advice. No system, model or process described here is claimed to produce any level of accuracy, profit or return, and none should be inferred. Regulatory positions stated here were current as of 18 July 2026 and should be verified at the source. Bharath Shiksha is an educational publisher, not a SEBI-registered investment adviser or research analyst.
Common Questions
Frequently Asked Questions
Can AI predict stock prices?
+No, not in any form a retail trader can rely on, and the reasons are structural rather than a matter of waiting for better models. Day-to-day price changes are dominated by noise, so the predictable part of a return is a thin tilt rather than a call you can act on with confidence. The process generating those returns also keeps changing as participants, rules and volatility regimes turn over, so a relationship fitted on the past is fitted to a world that has partly ended. On top of both, a flexible model searching a large space will always find something that looks convincing on the data it was given. A system that forecast prices well enough to trade would be among the most valuable objects in finance, and the fact that one is being offered cheaply is itself the tell.
Does machine learning work for trading at all?
+Yes, on narrow jobs, and that is a real answer rather than a consolation prize. Machine learning earns its keep when it augments a process you already understand: scoring how well an order was executed, classifying the broad state of the market, ranking a watchlist against criteria you wrote down, flagging trades in your own record that do not match your rules, and turning documents into structured fields you then verify. What those jobs share is that examples are available, the answer can be checked without waiting on the market, and a wrong answer costs time rather than money. Prediction shares none of those three properties, which is exactly why it is the one job the marketing sells.
Why does an AI strategy look brilliant in a backtest and fail live?
+Because a beautiful backtest is the ordinary output of a search, not evidence of an edge. If you try many variations of a rule and keep the best, the winner's record is the maximum of a distribution rather than a measurement of skill, and the maximum of enough noisy draws is always impressive. A simulation of two hundred rules built to have a true edge of exactly zero produced a best in-sample Sharpe of about 1.7 over two years, close to what the arithmetic of search predicts, and that same rule then ranked in the middle of the pack on data it had not seen. Model capacity does the same thing from the other direction: every additional input improves the fit on the data the model was given while degrading it on data it was not.
Can I use a general-purpose language assistant for trading?
+For study, code and text work, yes. For forecasts and strategy generation, no. A large language model is trained to continue text plausibly, so it is genuinely useful for explaining a concept, drafting or reviewing code, turning a vague idea into a written specification, and pulling named fields out of a long document that you then check against the source. Asking it what a price will do is a category error: it has no mechanism for that quantity, and it will answer fluently regardless, because fluency is what it was optimised for. The sharpest practical hazard is confident code with a subtle look-ahead bug, which is more dangerous than no code at all.
How much data does a strategy need before its record means anything?
+Far more than most people assume, and the arithmetic is simple enough to do on paper. For an annualised Sharpe ratio, the t-statistic of a track record is roughly the Sharpe multiplied by the square root of the number of years, so the years needed to reach a given level of evidence scale as the square of the ratio between that level and the true Sharpe. A genuine annual Sharpe of 0.5 takes about sixteen years to reach a t-statistic of two, and about thirty-six years to reach three. A Sharpe of 1.0 takes about four years and nine years respectively. This is why a two-year backtest can be perfectly honest and still carry almost no information.
Should I learn machine learning before learning to trade?
+No, and the order matters more than the effort. Machine learning applied without market knowledge produces confident nonsense: labels that quietly encode the future, tests run on data the model has already seen, and inputs that mean nothing. Someone who understands how an order is filled, what a corporate action does to a price series and why costs scale with turnover will catch errors that no amount of modelling skill would surface. Learn to trade a written process first, measure it honestly, and bring machine learning in later as a filter and a checking tool on top of something you already understand.
Is the use of AI in India's securities market regulated?
+Partly, and the position is still forming. Chapter IIIB of the SEBI (Intermediaries) Regulations, inserted in February 2025, makes a regulated intermediary solely responsible for the data, the output and the legal compliance of any AI or machine-learning tool it uses, whether it built the tool or bought it. A broader consultation paper on the responsible use of AI and machine learning was published in June 2025 and, as of 18 July 2026, had not been finalised into a binding framework. Separately, the retail algorithmic trading framework became applicable to all stock brokers on 1 April 2026. These rules move; confirm the current position at the source before you build anything on it.
Can a retail trader run AI-driven high-frequency strategies in India?
+No, and no amount of model sophistication changes it. High-frequency trading is an infrastructure business: colocation at the exchange, engineered low-latency systems, and exchange-approved strategies. The advantage is measured in microseconds and capital expenditure, not in cleverness, so a better model does not substitute for any of it. Retail machine learning, if it has a place at all, lives at daily or swing horizons where the constraint is process and discipline rather than latency, and where the useful jobs are the checking and classifying ones rather than the predicting one.
Does using a model reduce the need for risk control?
+No, and assuming it does is the most expensive mistake on this page. A model is a filter or an overlay, and it changes nothing about position sizing, stop placement or the arithmetic of expectancy that decides whether an account survives a bad run. Automating a process that has no edge simply produces the losses faster and more consistently. Because a model's output arrives with an air of objectivity, it can also make a trader size up on a signal they would have questioned if a person had suggested it, which makes the risk layer more important with a model in the loop rather than less.
Where the facts come from
Sources
- The expected best Sharpe ratio from a search. David H. Bailey, Jonathan M. Borwein, Marcos Lopez de Prado and Qiji Jim Zhu, Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance, Notices of the American Mathematical Society 61(5), 2014. Derives the Sharpe ratio the best of a given number of zero-edge trials will show over a given history, which is the calculation behind the fourth number in the first figure and behind the tool on this page. ams.org
- Correcting a Sharpe ratio for the breadth of the search. David H. Bailey and Marcos Lopez de Prado, The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting and Non-Normality, Journal of Portfolio Management 40(5), 2014. Establishes that the number of trials is part of the result and must be reported with it. papers.ssrn.com
- The multiple-testing hurdle. Campbell R. Harvey, Yan Liu and Heqing Zhu, on multiple testing in the cross-section of expected returns, Review of Financial Studies 29(1), 2016. Argues that given the number of candidate factors the literature has tested, a newly reported result should clear a t-statistic of roughly three rather than the conventional two. This is the upper curve in the second figure. papers.ssrn.com
- Out-of-sample failure of return predictors. Ivo Welch and Amit Goyal, A Comprehensive Look at the Empirical Performance of Equity Premium Prediction, Review of Financial Studies 21(4), 2008. Finds that variables which predict the equity premium well in sample generally fail to do so out of sample, which is the empirical backbone of the claim that the forecastable component of returns is thin. academic.oup.com
- Responsibility for AI and machine-learning tools in India. SEBI (Intermediaries) (Amendment) Regulations, 2025, notified 6 February 2025, inserting Chapter IIIB and Regulation 16C: a regulated person using AI or ML tools, built in house or procured, is solely responsible for data privacy, security and integrity, for the output, and for compliance with applicable law. Verified 19 July 2026. sebi.gov.in
- The proposed AI and ML guidelines, still at consultation stage. SEBI, Consultation Paper on Guidelines for Responsible Usage of AI/ML in Indian Securities Markets, 20 June 2025, comments closed 11 July 2025. Paragraph 5 lists five core guiding principles: model governance; investor protection and disclosure; testing framework; fairness and bias; and data privacy and cyber-security measures. Not finalised into a circular or regulation as of 18 July 2026. The widely repeated "six principles" list does not appear in the paper. sebi.gov.in
- When the retail algo framework took effect. SEBI circular SEBI/HO/MIRSD/MIRSD-PoD/P/CIR/2025/0000013 of 4 February 2025 on safer participation of retail investors in algorithmic trading, as extended by circular SEBI/HO/MIRSD/MIRSD-PoD/P/CIR/2025/132 of 30 September 2025, which set the framework as applicable to all stock brokers with effect from 1 April 2026. Earlier dates of 1 August 2025 and 1 October 2025 are superseded. sebi.gov.in
- Outcomes in the equity derivatives segment. SEBI press release of 23 September 2024 reporting that 93% of over one crore individual F&O traders incurred losses averaging about ₹2 lakh each, inclusive of transaction costs, over FY22 to FY24, with aggregate losses exceeding ₹1.8 lakh crore. The figure is a three-year cumulative measure; SEBI's follow-up study of 7 July 2025 reports about 91% of individual traders in net loss in FY25 with net losses of ₹1,05,603 crore. sebi.gov.in
- Data protection obligations and their timetable. Digital Personal Data Protection Rules, 2025, notified by the Ministry of Electronics and Information Technology in November 2025. Board provisions took effect on publication, consent-manager registration follows at twelve months, and the core obligations on data fiduciaries, including notice, consent, children's data and breach reporting, commence eighteen months after publication. Verified 19 July 2026. meity.gov.in
On the figures. All five are computed rather than drawn. Each is generated by a simulation whose parameters are stated in its caption, and the numbers reported alongside each figure are outputs of that simulation, not illustrative values chosen to make a point. Where a figure and a cited source disagree with a claim commonly made about this subject, the figure and the source are what this page reports.