From 1 Jul 2025 to 17 Jul 2026 the Nifty ended lower than it began, and in that time its day to day behaviour changed sign. In the quarter to 3 Oct 2025 a rise tended to be followed by a small rise. In the quarter to 16 Apr 2026 a rise tended to be followed by a small fall. Over the whole stretch the pull was close to nothing at all. Ask a momentum trader and a mean reversion trader to read the same file and each can find a quarter that proves them right. This lesson tests what the file supports with two plain measures, autocorrelation and the variance ratio, shows how little four quarters can tell you, and explains why the two styles are less opposed than they sound. Every figure comes from NSE end of day files, dated and not live.
3By the end you can
- Say how momentum and mean reversion differ by horizon and by object, a price or a spread.
- Read an autocorrelation and a variance ratio next to their noise bands.
- Explain why a short window cannot name a regime and what a fair test must fix first.
The idea, plainly
Two bets about the last move
Momentum says a price that has been rising tends to keep rising over the next holding period, and a falling one to keep falling. Mean reversion says a price stretched away from an average tends to come back. They are not opposite beliefs about markets in general. They are bets at different horizons and on different objects. Studies of many markets report momentum over months in prices, and reversal over days to a week for some stocks and indices. Mean reversion in the strict sense belongs to a spread between two prices that share a home, which the cointegration module tests. A single index level is much closer to a random walk. Run the same unit root test on the Nifty close and it cannot reject a wandering path, where the daily returns reject it easily.
Measuring the pull
Autocorrelation asks whether today's return is related to the return k days ago. A positive value points to momentum, a negative one to reversion. The variance ratio asks the same thing over longer horizons. If returns were unrelated, the variance of a q day return would be q times the variance of a one day return. A ratio above one says moves tend to run, and a ratio below one says they tend to reverse.
The maths: autocorrelation and the variance ratio
For daily log returns r, with the sample mean removed:
AC(k) = cov(r_t, r_t-k)
/ var(r)
VR(q) = var(q day sum)
/ (q x var(1 day))
With n returns and no true pull, AC(1) has a spread of about 1 over the square root of n. For 258 returns that is about 0.06, so a value inside about two spreads of zero cannot be told from noise. The variance ratio has a wider spread that grows with q, because overlapping q day sums share most of their days. Read a five day ratio below one next to that wider spread, not next to zero.
3 terms this section turns on
Why the options desk cares
The two styles are also two payoffs. Someone long options, such as a gamma scalper, gains when moves run beyond what the option price assumed. Someone selling options earns the premium when moves stay inside it, and loses when a trend carries the index past the strikes. So a regime question is a risk question for both sides. A seller who assumes reversion carries the risk of a sustained run, and a buyer who assumes momentum pays decay while the index ranges. The variance risk premium module puts a price on that trade.
Worked example: four quarters of the Nifty
The file gives 259 sessions from 1 Jul 2025 to 17 Jul 2026. Here is the path first. It has a flat stretch, a rise, a long fall and a rebound, so a story of any style can be told about some part of it.
One path with a flat stretch, a rise, a fall and a rebound
Autocorrelation by quarter
Split the returns into four blocks of about 65 sessions and measure the lag one autocorrelation in each.
The sign of the pull flipped once, inside the noise
Read one more paragraph, 66 words
Three quarters lean toward momentum and one toward reversion, and every bar sits inside the noise band of two spreads, about 0.25 either side of zero. The highest is a little past one spread. A regime change would need a shift larger than the band, held for longer than a quarter. Four blocks cannot show one. They show that a short window can flatter either style.
The variance ratio
The variance ratio looks at the same returns summed over several days. Below one means moves tended to reverse over that horizon.
All four ratios sit a little below one, which is within their noise
All four sit a little under one, a mild lean toward reversal. The five day ratio is the lowest, and it is still only about one spread below one. This is the honest answer to the question this lesson began with: a faint reversal tilt that this sample cannot separate from no pull at all.
Two simple rules, counted
A rule turns the idea into a count. The first rule looks at the last 20 sessions. If the index rose, momentum says the next five sessions rise. The second looks at how far the close sits from its 20 session average, in usual sizes. If it sits more than one and a half above, reversion says the next five sessions fall. Nothing is traded and no return is claimed. The question is only how often the next five sessions agreed.
The two rules agreed with the next week about as often as a coin would
Read one more paragraph, 104 words
After a 20 session rise, the next five sessions rose less than half the time, which leans against momentum. After a stretch above the average they fell about two times in three, which leans for reversion. After a stretch below, they rose only two times in five, which leans against it. The two stretch cases pull in opposite directions, and put together they agree with the rule on about half the cases. The windows overlap, so only about a fifth of the cases are independent looks. A sample this size and this shape is a demonstration of method. It does not choose a style.
What a fair test must fix first
Each of the choices above could have been made differently, and each changes the count. A fair test writes them down before it looks: the lookback, the holding period, the threshold for a stretch, the data used to choose them and the data used to judge them. It also includes costs. A reversion rule trades often, because it acts on every stretch. Every round trip of a futures leg pays STT of 0.05% on the sale and the other statutory charges in the cost module, so a small average gain per trade can vanish. A momentum rule trades less but holds through the fall that ends a trend. Neither cost shows in a count of ups and downs. The same goes for slippage, the gap between the price you see and the price you get, which grows when a rule fires on exactly the days when everyone else is moving too. Add it to the cost before judging an edge that is smaller than a fraction of a percent per trade.
Read one more paragraph, 67 words
Published studies add that the effects differ by market, period and horizon, and that many fade once costs are included. Indian studies exist, and the one paper in the sources reports momentum across many futures markets over months. None of that is a result for the Nifty on these dates. The only claim this lesson tests is the narrow one in its charts, on the dates shown.
Read the chart yourself
Look at the path above and find the longest stretch where the index kept moving one way. Ask what lookback and holding period would have called it momentum, and which would have called its end a reversal. If two choices of the same rule give opposite answers on the same stretch, the rule has a free parameter that the data did not fix. That parameter should be written down before the test, and the test run on a period that did not shape it.
The common mistake
The claim: Indian markets are mean reverting, so every sharp move gets reversed within the week.
The sample points the other way as often as it points this way. The five day variance ratio was a little below one, which fits the claim, yet it sat within one spread of no pull. Three quarters had a positive autocorrelation, which fits the opposite claim. After a stretch below the average the next week rose only two times in five. A single regime label, either way, needs far more data than four quarters. Ask for the lookback, the holding period and the threshold before the test, the count of cases including the misses, and a result that survives on a period that was not used to build it.
Key takeaways
- Momentum and reversion are bets at different horizons, not opposite worldviews.
- A single index is near a random walk; a spread can have a home.
- Read autocorrelation and variance ratios against their noise bands.
- Four quarters can flatter either style without proving it.
- Fix the lookback, holding period and threshold before the test.
Two terms to hold on to from 18.9
Before you trade. Derivatives trading carries a high risk of loss. SEBI's study "Profitability of Individual Traders in the Equity Derivatives Segment (FY25 to FY26)" (August 2026) found that 87.7% of individual traders in equity F&O made net losses in FY26, with aggregate net losses of Rs 91,685 crore. Read the SEBI study