Research · counted from the record
Mean reversion in Indian stocks: right 82% of the time, for almost nothing
A popular mean-reversion rule, run on 210 Indian F&O stocks: its signals came back to the average 82% of the time. Its average result per signal, before costs, was +0.14%. Being right four times in five was worth almost nothing.
The idea
Mean reversion is the belief that a price stretched far from its recent average tends to come back. It is one of the oldest ideas in trading, and screeners built on it are everywhere.
The appeal is the hit rate. Stretched prices do come back, most of the time. The question worth asking is what the times they do not come back cost.
How it was tested
The test used daily closes of 210 stocks that trade in NSE's F&O segment, from 9 June 2021 to 9 October 2026, with splits, bonuses and demergers adjusted out.
The rule. Take an 8-day average of the close. Measure the usual size of a day's move as the standard deviation of daily point changes over the previous 756 sessions (about three years). A stock is stretched when it closes more than 1.5 of those away from its average, and extreme beyond 2.5.
The signal is the close coming back inside the 1.5 band. It is followed until the price reaches the average (a win for the idea), goes beyond the extreme band the wrong way (a stop), or ten sessions pass.
Everything is measured as it stood on each day, with no later data: the average, the spread and the bands are the ones a trader would have seen that evening.
What happened
| The original rule (8-day average) | Bollinger bands (20-day, 2 SD) | |
|---|---|---|
| Finished signals | 9,144 | 11,682 |
| Reached the average | 82.3% | 48.7% |
| Stopped out beyond the extreme band | 14.7% | 2.7% |
| Ran out of time (ten sessions) | 3.0% | 48.6% |
| Ended in profit | 66.7% | 57.1% |
| Average result per signal | +0.14% | −0.01% |
| A typical win | +1.63% | +3.48% |
| A typical loss | −3.05% | −4.17% |
| Typical days to reach the average | 2 | 5 |
| Worst single signal | −23.2% | −49.0% |
Why a high hit rate earned so little
The wins were small and the losses were large. A typical winning signal made about 1.6%; a typical losing one lost about 3.0%, nearly twice as much.
That is the shape of mean reversion. The price usually drifts back a little, and now and then it does not come back at all: it keeps going, because something real has changed. Those few cases pay for all the small wins.
An average of +0.14% per signal is before brokerage, securities transaction tax, exchange charges and the gap between the price on the screen and the price one actually gets. After those, it is below zero.
The wider Bollinger version did no better. Its bands are further out, so its signals are rarer per stock and its wins bigger, but nearly half never reached the average in time, and the average result was nil.
What it does not show
- It does not show that mean reversion never works. It shows that this plain rule, applied to every F&O stock alike, has no edge to speak of. Filters for trend, liquidity or news might change that, and each would need its own test.
- The signals overlap in time: on a day the whole market falls, dozens fire together. The count of signals overstates the number of separate bets.
- Five years is one stretch of market history, most of it rising.
Questions people ask
Does mean reversion work in Indian stocks?
A plain version tested on 210 F&O stocks over about five years reached its target 82% of the time but averaged only +0.14% per signal before costs, because losses were about twice the size of wins.
Is a high win rate enough for a trading strategy?
No. What matters is the win rate together with the size of wins and losses. A rule that wins 82% of the time and loses twice as much when wrong can still earn nothing.
Do Bollinger Bands work on NSE stocks?
Used as a plain reversion rule (20-day average, 2 standard deviations), 49% of signals reached the average within ten sessions and the average result was −0.01% before costs.
Can I change the formula and test it myself?
Yes. Schema's mean-reversion page lets a reader change the average, the measure of spread, the bands and the exit, and recounts the record for every stock at once.
Keep reading
Research and information, not investment advice. Every figure describes the past: it is what happened, counted, and not a forecast of what will.