Financial engineering for promos, bonuses and gamification

When you send a customer a $20 code, you have not spent $20. You have written a contract that pays $20 if the customer meets a condition before a date, and pays nothing otherwise. There is a strike, which is the qualifying purchase. There is an expiry, which is the end of the promo window. There is a holder who exercises only when it suits them, and a writer who is obliged to honour it either way.

That is an option, and you are short it. The distinction matters because the two positions are accounted for differently in every other part of a business. Nobody books a written call at its notional value, because the notional is not what it costs. What it costs is the probability-weighted payout, and that is a number you have to model rather than read off the offer.

Marketing teams do the opposite. They budget promos at notional (audience size times face value), discover the real payout in the following month's invoice, and call the difference a forecasting problem. It is not a forecasting problem. It is a pricing problem that was never treated as one.

The consequences compound as the mechanic gets more complex. A flat discount is a simple option and the error stays small. A bonus with a wagering requirement is a path-dependent one, where the payout depends on the entire sequence of a player's balance rather than its endpoint, and intuition stops working completely. Simulated across 200,000 players, a $100 bonus with a 30x rollover costs $52.63 rather than $100, because only 52.63% of players survive the turnover requirement to realise it. A team that budgets that bonus at face value over-reserves by 90%. A team that budgets it at zero because "most people never clear it" is wrong in the other direction and finds out in a quarter where variance runs hot.

Streaks, leaderboards and loyalty tiers are the same problem wearing product clothing. A streak is a sequence of small options with a rising exercise probability, because a user on day 20 is a different risk than a user on day 1. A leaderboard is a tournament, and tournament theory has known since the 1980s that effort collapses once a competitor's win probability approaches zero. On a single global leaderboard of a thousand players with ten prizes, only 5% of the field has a realistic shot. The other 95% are being shown, every day, a ranked list confirming they cannot win. Split the same players into brackets of ten and 48.7% of them have a real chance, with no change to the prize budget at all.

A loyalty tier is the case where the intuition is most confidently wrong. The stated logic is that thresholds make people spend more to qualify, and they do. Modelled across 200,000 customers with a realistic spend distribution, threshold stretching generates real incremental margin at every threshold tested. It also never once pays for the perk, because the perk goes to everyone above the line while the stretching only happens just below it. At a $300 threshold the stretch is worth $157,924 and the perk costs $1,164,870. The tier is not a growth mechanic at that point, it is a discount with extra steps, and the only thing that can rescue it is a retention lift on qualifiers of at least 6.79%. That number is knowable before launch. Almost nobody calculates it.

None of this requires a quant team or a trading desk. The models on this site run in Python with NumPy and the standard library, and the longest is under a hundred lines. What it requires is treating the offer as an instrument to be priced rather than a creative decision to be approved, and running the numbers while you can still change the terms.