Guide · Research method · Updated 2026-09-19

The sample size that would settle the question is larger than the run in which you would quit

Every number on this page was computed rather than quoted, and the method is stated so it can be redone. That matters more than usual here, because this is the one topic where a confidently asserted figure does the most damage.

The question is badly posed, and fixing it is most of the work

"How many trades do I need?" has no answer, in the same way that "how far away can I read a sign?" has no answer until somebody says how large the letters are. The missing quantity is the effect size: the average result per trade divided by the standard deviation of results per trade.

Expressing it that way is not a technicality. It strips out position size and account size, which are the two things people usually reach for and which have no bearing on whether an edge is detectable. A system that averages a fifth of its own variability per trade is easy to establish. A system that averages a fiftieth of its own variability is nearly impossible to establish, no matter how much money passes through it.

Once the question is posed properly the arithmetic is short. You are asking for the sample at which a test would reject the hypothesis of no edge, at a chosen significance, with a chosen probability of actually doing so if the edge is there. That second quantity is power, and leaving it unspecified is how most published figures end up far too small.

The wall

Read across any row and the cost of being more careful is visible. Read down any column and the quadratic term is visible. Between an edge of 0.10 and an edge of 0.05, the edge halves and the requirement goes from 619 to 2,474, which is four times.

This is the structural fact that makes the whole subject difficult. Large edges are rare and do not survive long once found. Small edges are the realistic case, and small edges sit on the steep part of this curve, where the data requirement runs away from you.

Converting trades into years

A trade count is abstract until it is put on a calendar. In 2026 the Indian equity market has 261 weekdays and 15 weekday trading holidays, which leaves 246 trading days. That is the divisor.

The temptation on seeing this table is to read along the bottom row and conclude that frequency solves the problem. It does not, for a reason that is easy to state and easy to forget: raising frequency changes the system being measured. Costs rise, the holding period shortens, and the edge per trade usually falls. Since the requirement is quadratic in that edge, a change that halves it multiplies the target by four while multiplying your rate of accumulation by rather less. The target generally moves away faster than you move toward it.

Trade count is not sample size

The arithmetic above treats each trade as an independent observation. Real trading records are not like that. A system that reads a trend holds broadly the same view for as long as the trend lasts, so a run of results reflects one market condition rather than many independent draws. When consecutive results are correlated, each additional trade contributes less than one observation's worth of evidence.

Modest correlation does real damage. A correlation of 0.20 between consecutive results, which is unremarkable for a system holding a directional view, turns 2,474 trades into about 1,649 observations. The record has not changed. What it can support has.

This is also why pooling years of results into one figure flatters the analysis. The pooled count is large, and the independent content in it is a good deal smaller than the count suggests.

The asymmetry that decides the matter

So far this is a data problem, and data problems feel solvable with patience. The next result is the one that changes the shape of the argument.

Take a system whose edge is real. Not assumed, not hoped for: real by construction, set at 0.05 standard deviations per trade, which is the row of the first table requiring 2,474 trades. Simulate 20,000 traders running that system for exactly that many trades, and ask a different question from the usual one. Not "do they end up ahead", but "do they get there at all", where getting there means never having passed through a peak-to-trough decline large enough that an ordinary person would have stopped.

At a stopping point of twenty standard deviations, 91.7 per cent of these runs end early. Every one of those traders had a genuine edge. None of them reached the point at which they could have shown it, and each of them experienced a decline that looked exactly like evidence the system did not work.

That is the asymmetry. The sample required to confirm a small edge is larger than the sample within which the ordinary variation of that same edge will produce a run bad enough to abandon it. The two quantities are not close. Nothing about this is a claim that any particular system works or does not. It is a statement about what a small edge feels like from the inside while it is working.

What a forward test can actually settle

The standard advice when a backtest looks good is to forward test it before committing, usually for a stated period such as three months. That advice is sound in its intent and almost always misapplied, because the period is chosen by how long a person is willing to wait rather than by what the period can resolve. The arithmetic runs the other way just as easily: fix the number of trades, and it tells you the smallest edge that trial could have detected.

A three month forward test at three trades a day accumulates about 184 trades, and could only have detected an edge of about 0.18 standard deviations per trade. Anything smaller than that passes through such a test invisibly, in either direction. A genuinely good system and a worthless one both produce a three month record that is consistent with the other explanation, which means the test did not distinguish them and the confidence gained from passing it was not earned.

There is a sharper way to put it. A forward test of that length is capable of rejecting the claim that a system has a large edge. It is not capable of establishing that it has a small one. Since almost every real edge is small, the test is nearly always being used for the thing it cannot do. That is not an argument against forward testing, which catches implementation faults, cost errors and outright broken logic that no amount of historical data will reveal. It is an argument against reading survival through one as evidence about the edge.

What the population statistics can and cannot tell you

In August 2026 the market regulator published a fresh pair of studies on individual traders in the equity derivatives segment covering the two years to FY26, built on client-level and transaction-level data covering roughly nine tenths of individual participants in the segment. The headline figures were that about 87.7 per cent of individual traders were net losers in FY26, that aggregate net losses fell by roughly 18 per cent to about Rs 91,685 crore, that participation fell by about 18 per cent to roughly 87.5 lakh individuals, and that options accounted for the large majority of the aggregate losses.

Most published commentary still works from the earlier study covering FY22 to FY24, which is now superseded on both the period and the numbers. Checking which study a figure came from is worth doing before repeating it.

The more important point is what a figure of that kind can support. It is a population result, and it is strong evidence about the population: at that scale, the aggregate outcome is not noise. What it cannot do is separate two very different groups inside the losing majority. One group had no edge. The other had one and stopped inside the ordinary variation of it, exactly as the simulation above describes. Nothing in a population loss rate distinguishes them, and the arithmetic on this page says the second group is not a rounding error.

What to do instead of waiting for a sample that will not arrive

If confirmation is out of reach, the useful move is to change what the data is for. A few hundred trades cannot certify an edge. They can do four other things, and each of them is answerable now.

They can eliminate. A claim of a large edge would have shown up in a small sample, so a small sample is enough to reject it. Most of what a researcher needs to discard is discarded this way.

They can detect breakage. Asking whether recent results are consistent with the historical distribution is a much easier question than asking whether that distribution has a positive mean, and it is the question that matters when something has stopped working.

They can verify the cost model. Costs and slippage are observed directly rather than inferred, so they converge on far less data than an edge does. A cost model that is wrong invalidates every edge estimate built on it, and it is checkable this month.

They can record the trial count. The single most valuable thing a researcher can keep is the number of variants tried and discarded, because that number is what makes any later significance claim honest, and nobody reconstructs it afterwards.

What remains after that is judgement: reasoning about why a mechanism should work, whether the conditions that produced it still hold, and what would have to be true for it to stop. That is not a consolation prize for lacking data. Given the arithmetic above, it is the only tool available at the sample sizes a person actually has, and treating it as a discipline with its own standards is the difference between a process and a hope.