A social media sentiment signal on 150 consumer stocks goes long the top decile and short the bottom, with a 53% hit rate on two-day holds and an average move of 1.2%. Costs are 15 basis points a side and capacity is capped at 1% of daily volume. Is there a business, and how big?

Not as it stands: the signal earns about 7.2 basis points a trade before costs and pays 30 to trade, losing about 23 basis points a trade. It would need a 62.5% hit rate at these costs. Capacity is also small, about Rs 4.5 crore of gross positions. The business, if any, is as an overlay on trades already being made, or with a longer hold if the edge persists.

Think of a coin that lands heads 53 times in 100 and pays Rs 120 either way. You collect 53 x 120 and pay 47 x 120, a net of Rs 720 over 100 tosses, Rs 7.20 a toss. The edge is the hit rate's excess over one half, doubled, times the average move: 0.06 x 1.2%, which is 7.2 basis points a trade. A 53% hit rate sounds like a strong signal; on symmetric moves it is a thin one. It also assumes winners and losers are the same size, which you would check in the data before anything else.

Reverse the arithmetic. To cover a 30 basis point round trip on 1.2% moves, the hit rate must satisfy (2p - 1) x 120 = 30, which is p = 62.5%. Or, at 53%, the round trip must cost under 7.2 basis points, which no one executing a two-day consumer stock rotation achieves. The two honest routes are a longer hold, if the signal keeps predicting beyond two days, so that one round trip pays for a larger move, and trading it where the cost is already paid.

The backtest has about 11,250 trades, and a naive standard error on a hit rate is the square root of 0.25 over that count, 0.47 percentage points, which makes 53% look like 6.4 standard errors. But the 30 trades placed on the same day share one market mood, so they are closer to one bet than thirty; counting one independent bet per holding period gives 375 bets and a standard error of 2.6 points, and 53% is only 1.2 standard errors from a coin. The truth sits between the two, and it needs checking with a test that clusters trades by date. Add that social media data invites many tried variants, and the 53% may be the best of several, which the multiple testingTrying many versions of a signal and keeping the best one. The winner looks stronger than it is because some of its edge is luck. problem shrinks further.

At 1% of a Rs 15 crore daily volume, each position can be about Rs 15 lakh, so 30 names hold about Rs 4.5 crore of gross positions. Even if the costs vanished, 7.2 basis points on that book turning over every two days is about Rs 41 lakh a year, too small to pay for the data and the people on its own. The business is as an input: a fund already trading these stocks for other reasons can lean each order by the sentiment score and collect a few basis points on trades whose costs it pays anyway. That is how most thin signals earn their keep.