The Factor Zoo Doesn't Travel: What Breaks When You Run Global Quant on Indian Equities

A factor is a portfolio, not a fact

Start with the model everyone is implicitly fitting. The return of stock i over a period decomposes into a part explained by its exposures to a handful of common factors and a part that isn't:

A "factor" in the trading sense is a portfolio engineered to load on exactly one fk — long the stocks high on the characteristic, short the ones low on it, with the other exposures netted out. The "premium" everyone quotes is just the time-series mean of that portfolio's return. Momentum is the long-winners / short-losers portfolio; low-vol is long-low-beta / short-high-beta; and so on.

The headline statistic for such a portfolio is its information ratio, and the cleanest way to think about where it comes from is Grinold's fundamental law of active management:

where IC is the information coefficient — the cross-sectional correlation between your forecast and the realised return — and BR is breadth, the number of genuinely independent bets you take per year. A weak signal applied to a thousand names can beat a strong signal applied to ten. This is the number the papers report, and it is entirely a gross number. It tells you how good the signal is. It tells you nothing about whether you can keep any of it.

The only equation that decides anything

What you actually keep is gross alpha minus what it costs to run the portfolio, and the cost scales with how often you trade:

Here T is annual turnover — how much of the book you replace in a year, counting both sides — and c is the all-in cost of moving a unit of notional. Set αnet to zero and you get the only threshold that matters:

A factor is investable only if its turnover sits below its breakeven turnover T*. Above that line, a beautiful gross IR delivers nothing, or less than nothing. The literature publishes αgross. Your prime broker and your exchange publish c. The entire question of whether a famous factor travels to a new market is the question of what happens to that ratio when you swap in the local cost stack. So let's build the Indian one.

The cost stack on the NSE

Decompose c into its parts:

The half-spread and the exchange/SEBI/stamp fees are small and familiar. Two terms are not, and they are the two that a US-calibrated model gets badly wrong.

The first is the Securities Transaction Tax. On delivery-based equity it runs on the order of 10 basis points a side (the precise figure moves with each budget, so treat the number as illustrative — the structure is the point). Crucially it is a tax on turnover, not on profit: you pay it whether the trade made money or not. A factor that turns its book over five times a year is paying something like 100 bps a year in STT alone, before a single rupee of spread or impact. When the gross premium you are chasing is itself only 200 to 400 bps, that is not a rounding error. It is half the edge, gone, to a line item that never appears in a US factor backtest because no such tax exists there.

The second is market impact, and the standard model is the square-root law:

where Q is your order size, V is the stock's average daily volume, σ its daily volatility, and Y a constant of order one. The trap in India is the V in the denominator. Headline market capitalisation badly overstates the tradeable size of an Indian listed company, because promoter holding is high — for a large share of the market, half or more of the shares are locked in founder, family, or group hands and never circulate. The float that actually trades is a fraction of the cap. Two stocks with identical market caps can have free floats that differ by three times, which means identical-looking positions generate wildly different impact. Calibrate V off market cap the way a developed-market model implicitly does and you will systematically under-price the cost of every trade in the names where promoters sit tightest — which tend to be exactly the smaller, higher-expected-return names a factor wants to own.

Momentum: the factor that pays for its own funeral

Cross-sectional momentum is the cleanest case. Rank the universe by trailing twelve-month return, skipping the most recent month to sidestep short-term reversal, go long the top decile and short the bottom. Of the standard factors it tends to post the highest IC. It also has the highest turnover by a wide margin — winners and losers reshuffle every month — which means it sits furthest above its breakeven T*. The gross premium is real. On the NSE most of it dies in the T · c term, where high turnover runs straight into the STT floor and the float-adjusted impact term compounds it.

That is before you get to the tail. Momentum's payoff is option-like and negatively skewed: long stretches of steady gains punctuated by violent crashes. The reason is that its market beta is not constant. After a deep drawdown the short leg is full of beaten-down, high-beta names, so the portfolio quietly accumulates large negative beta — and when the market rebounds it gets run over. This is what Daniel and Moskowitz labelled momentum crashes, and structurally the strategy behaves like writing a put on the market: you collect premium for years and pay it back in a week.

India adds its own twist to the tail. Individual scrips trade inside daily price bands — commonly ±2, 5, 10, or 20 per cent depending on the security. The tails that momentum lives off are precisely the explosive moves the band truncates. A stock locked limit-up is a signal you can observe and a fill you cannot get; the return series your backtest learns from is censored at the band, and a naive test happily books the limit-up close as a tradeable price. Censoring a fat-tailed distribution biases the measured IC upward — which means the factor most flattered by the bad data is the one whose live performance disappoints hardest.

Low volatility: the one that travels better

Now the factor that survives the trip, and even gains on it. The CAPM predicts that higher beta earns higher return. Empirically the security market line is far too flat: low-beta stocks deliver more return per unit of risk than high-beta ones. Frazzini and Pedersen's explanation is leverage. An investor who wants more return but cannot or will not borrow does the only thing left — overweights high-beta stocks as a poor man's leverage. That bids high beta up and leaves low beta cheap. The betting-against-beta portfolio harvests the slope error directly: go long low-beta names levered up to a beta of one, short high-beta names de-levered to a beta of one, and you hold a market-neutral book that pays as the line re-flattens.

Two things make this factor behave well on the NSE. First, the marginal participant is unusually leverage-hungry and reaches for that leverage through the derivatives market and through high-beta single names rather than through margin on a sober low-beta book — which pushes high beta even further from fair value and widens the very distortion the factor monetises. Second, and decisively, it is a low-turnover factor. Betas are persistent quarter to quarter, so you rebalance slowly, you sit comfortably below T*, and the STT floor barely grazes you. You do not need a high IC when your turnover is low — the fundamental law gives you the gross edge and the cost stack lets you keep most of it.

The constraints that never make it into the backtest

Costs shrink the premium. Constraints remove trades you assumed you could make.

The F&O ban period. When the aggregate open interest in a stock's derivatives crosses 95 per cent of its market-wide position limit, the exchange bars new positions until open interest falls back — only trades that reduce existing positions are allowed. Your feasible set contracts exactly when a name is most crowded, which is frequently exactly when your signal on it is strongest. The backtest assumes you can always put the trade on. The live book finds the door locked on its best ideas.

Borrow and the short leg. A clean long-short factor assumes you can locate and finance a short. Single-stock borrow in India is thin and expensive outside the liquid F&O names, so the elegant market-neutral portfolio in the paper is, in practice, often a long-only tilt plus an index-futures hedge. That is a different object with a different risk profile — and the asymmetry quietly truncates whichever side of every factor lives on the short leg.

Censoring, again. Price bands, trading halts, and circuit breakers all do the same statistical violence: they remove the observations from the extreme of the distribution, and they remove them non-randomly. Any IC, any Sharpe, any drawdown estimated on censored data is measuring a market that does not exist — a sanitised one where the worst and best days were clipped off before you looked.

Putting it together: in a high-cost market, slow factors win

Stack the net-alpha equation on top of the cost stack and an ordering falls out that the gross numbers would never give you. The factor that wins is not the one with the highest IC. It is the one that maximises

Momentum brings the IC and hands it back through turnover times (STT plus impact plus censoring bias). Low volatility and quality bring a more modest gross edge but turn over slowly, sit far below breakeven, and therefore keep most of what they earn. Value sits in between, dragged down by the borrow problem on its short leg. The raw ranking of factors by published Sharpe gets substantially reshuffled — sometimes reversed — once you divide through by the local T · c.

That generalises into a rule worth carrying to any market, not just India: the higher a market's cost-and-constraint load, the more the optimal book tilts toward low-turnover, capacity-light, slow-signal factors, regardless of which animal in the zoo has the prettiest gross backtest. The factor zoo is real. It simply does not travel at par. Every specimen has to be re-underwritten against the local microstructure — the tax on turnover, the float behind the impact, the bands that censor the tails, the ban-period that locks the door — before its number means anything where you actually trade.

I have spent a decade pricing the gap between a model's fill and a real one, and the lesson is always the same. The interesting work in Indian equities is not discovering a new factor. It is the unglamorous accounting — STT, float-adjusted impact, censored tails, ban-period feasibility — that tells you which of the famous ones you are actually allowed to keep. A backtest that ignores the cost stack is not being optimistic. It is quietly measuring a different market.