Where the edge stops

An intraday signal on Indian equities, measured across 150 symbols and 45,611 setups. Not "does it make money" — that question is too easy to answer wrongly. Instead: how far does price move against you before it moves for you, and how much is left once it does. The answers rule out the design the strategy started with.

Why measure the path and not the profit

Any intraday rule can be made profitable on a year of data by choosing the stop and the target after the fact. There are only two free parameters, a few thousand trades, and no shortage of combinations. An equity curve produced that way tells you what the search found, not what the market did.

So this study does not fit a stop or a target at all. It takes the entries the signal generates and measures the path that follows each one: how deep the adverse excursion goes before any target is reached, how often each target is reached, and what happens after it is. Those three quantities are properties of the data. A stop-and-target design is then something you either can or cannot build on top of them — and for this signal, the design it started with is one you cannot.

The signal

A volume-confirmed reclaim of the intraday VWAP, in the direction of the short-term trend. Signals are evaluated on the 5-minute close; the position is taken on the open of the next 1-minute bar, so no decision uses a price that had not printed when the decision was made.

entry when, on a 5-minute close:
    close > VWAP + 0.05 × ATR₁₅      reclaim, with a buffer scaled to volatility
    close > EMA20                    trend filter
    volume ≥ 10-bar average         participation filter

R      = ATR₁₅ at the signal bar     one unit of risk, per symbol, per moment
entry  = open of the next 1-minute bar
window = 10:05 to 14:00 entries · measured to 15:10 · one position at a time
reset  = after a trade, price must close back below VWAP before another is allowed

Risk is defined in volatility units rather than rupees, so a stop of "1R" means one 15-minute ATR — the same statistical distance on a quiet symbol and a violent one. Every number on this page is in those units. The sample is a full year of 1-minute data on 150 NSE names, restricted to symbols with at least 20 setups.

Finding 1 — the stop cannot go where the textbook puts it

The conventional construction is to risk 1R to make 2R. That requires the trade to survive without going 1R against you. So: how far does price actually travel against the entry before it first touches its 1R target?

The median symbol's median trade goes 1.21R against the entry first. 142 of 150 symbols have a median adverse excursion greater than 1R. At the 80th percentile the figure is 2.39R, and at the 90th, 3.19R.

This is the kind of result that only appears if you measure the path. A backtest with a 1R stop would have reported a poor strategy and a plausible-looking equity curve, and the conclusion would have been "the signal is bad" rather than "the risk model is wrong for the signal."

Finding 2 — reach decays quickly

Next: how often does a setup reach a given target at all? Every entry is swept against targets from 1R upward, checking only whether the level is touched before the session's cutoff.

About half of all setups touch 1R. By 1.6R it is under a third. The decay is smooth, with no shelf where a larger target becomes disproportionately worth taking.

The chart stops at 1.6R on purpose. Past that, the sweep starts exhausting symbols — it terminates for a symbol once reach falls below half its own 1R rate. Continuing the median past that point would measure which symbols survived, not what the population does. Where each symbol's sweep died is itself the answer to how far this signal carries:

Nearly three quarters of the universe runs out of measurable reach by 1.9R. That is the ceiling — the point past which this signal has nothing left to sell, on any symbol, regardless of how the position is managed.

Finding 3 — past the target, the path has no direction

The remaining hope for a strategy that cannot use a 1R stop is to hold: take a smaller, likelier target and let the good trades run. That requires the path after the target to be biased upward. It is not.

Measured from the moment a target is touched — how far price then retraces, against how much further it advances:

The two are the same size at every target level tested. The ratio of further gain to give-back stays between 1.05 and 1.08 from 1R through 1.6R: a few percent of asymmetry, on medians drawn from tens of thousands of setups, which is not a foundation for a trailing rule.

The edge, whatever there is of it, is entirely in getting to the level. None of it is in what happens afterwards.

Together the three findings close the design space from both ends. The stop cannot be tight, because ordinary noise exceeds 1R. The target cannot be far, because reach collapses by 2R. And the position cannot be held past its target, because the path beyond it is directionless. What remains is a narrow band, and a narrow band is the same thing as a small capacity.

The one forward simulation

Separately from the path study, a single symbol was simulated end to end to confirm the machinery works: setup detection, entry, stop, target, exit. ICICIBANK, 70 trades, December 2025 to April 2026, a 2R target against a 1R stop, and — when both are touched inside the same candle — the stop assumed to fill first. Timeouts exit at the close of the twelfth candle.

It ends positive: +6.55R over 70 trades, a mean of +0.10R per trade on a 28.6% win rate, which is what a 2:1 payoff buys you. It is also the least interesting number on this page, and the two figures beside it say why. The curve peaked at +14.63R and closed at +6.55R — it gave back 8.08R, 55% of its own peak, inside four months. A result whose drawdown is over half its high-water mark, on one symbol and 70 trades, is not evidence of an edge. It is evidence that the code runs.

What the filters showed, and why it does not count

Restricting to the middle confidence band produced a better result on fewer trades: 23 trades returning +11.43R, against +6.55R from all 70. Adding the remaining high-confidence setups back at 30% size gave +11.53R from 52 trades — nearly identical, from twice the activity.

That reads like a working filter, and it is not. The bands were drawn after the outcomes were known, on a sample of 70 trades from one symbol. With that many degrees of freedom a subset that doubles the return exists in almost any series, including a random one. The honest statement is that no confidence filter has been tested on data that was not used to choose it.

What this study does not establish

The last three matter. A walk-forward harness exists in the project and the variant engines run through it with round-trip costs and slippage modelled, but no fold-level results were retained, so this note presents no walk-forward equity curve — and a study that has not been validated out of sample should not describe itself as one that has. The path statistics stand on their own precisely because they fit nothing: there are no parameters in them to overfit.

Iterations that did not survive

Three variants were built and dropped, each recorded as its own script rather than edited away: a second afternoon entry window (14:00–14:45), a re-entry after a losing reclaim under stricter volume and strength gates, and a symbol-universe expansion with a conditional threshold relaxation. A separate comparison isolated the volume gate on its own across thirteen symbols. None of them changed the conclusions above, which is why the conclusions above are stated in terms of the path rather than any engine.

The data

The path study output is the full per-symbol sweep the figures are drawn from. The simulation files are the trade-by-trade record, including the losses.