How would you make money from social media data on Fizzora Beverages? You have three years of daily brand mentions and sentiment, and quarterly sales. Design the signal, the test and the trade, and say how you would know it is not noise.

Trade the part of mention growth that consensus does not already expect, and test whether it predicts the sales surprise, not sales. Raw mentions track sales at 0.91 but the surprise at only 0.21, because analysts see the same trends. The unexpected part correlates 0.75 with the surprise, but on twelve quarters, fitted in-sample. It is worth trading only after it holds out of sample, ideally across many brands.

The surprise, because that is what moves the price. If everyone in a housing society already knows the monsoon will be heavy, the umbrella seller's good season is no news to his landlord. Fizzora's price already reflects consensus sales growth, which itself correlates 0.97 with actual growth, so a signal that predicts sales growth mostly repeats what the market knows. Raw mention growth correlates 0.91 with sales growth but only 0.21 with the sales surprise to consensus.

Strip out what consensus already implies. Regress mention growth on consensus growth across quarters; the leftover, the residualThe part of a variable that a regression does not explain: here, the mention growth left over after allowing for what consensus already expects., is the part of the buzz analysts are not counting. Here mentions rise about 1.6 points for each point of consensus growth, and the residual correlates 0.75 with the surprise. Add sentiment as a second input only if it adds to that, and fix every choice, windows, weights and filters, before looking at the results.

Count your data points honestly first. With twelve quarters, a correlation needs to exceed about 0.58 to pass a 5% test, and that test assumes you did not choose the signal after looking at those same twelve points. The residual here was fitted on the same data, so 0.75 is an optimistic number. The fix is breadth: build the identical signal for forty consumer brands with the same data, 480 brand-quarters, where a correlation above about 0.09 is significant, and test it on quarters and brands that played no part in designing it. Check it with a placebo too: shuffle the dates and confirm the relationship disappears.

Across many brands, go long those whose unexpected mention growth is highest and short those where it is lowest, a few weeks before each results date, and close after results. That makes the trade about relative surprises and removes the market's direction. Two practical checks close the answer: whether the vendor's history was recorded at the time or reconstructed later, since backfilled data can quietly include information from the future, and whether the edge survives trading costs on the names that are liquid enough to trade.