The hedge desk.

The rule: train, then test

Every study split its history in two. Rules were chosen on 2020 through mid-2025 and then scored, untouched, on mid-2025 to now. A rule only counted if it made money after costs on both halves. That one discipline killed most of what looked promising.

It also exposed the trap that sinks most backtests. One sweep tried 1,463 rule combinations on stocks that suddenly came into play. Thirty-three of them cleared the bar on both halves. That sounds like a result until you remember that with 1,463 tries, a few dozen will look good by luck alone. The honest count of real edges in that sweep is somewhere between zero and a handful.

What survived, and what didn't

Buying the dip, modestly

The one family that held up everywhere: buy a stock that has fallen below its 20-day average and sell the first time it closes back above it. About +1% per trade after costs, on both halves, across 100,000+ episodes. Deeper dips pay more per trade and happen far less often. It is real and it is small, roughly a tenth of a percent per day of capital tied up.

Momentum flips with the regime

Hot industries and stocks already up 50% looked terrible on the training years and excellent on the test months. That isn't an edge; it's the 2025 market. A rule that changes sign with the regime is a bet on the regime.

The news is already in the price

A spike in articles about a stock predicted nothing useful at any holding period. Neither did buying straddles: the options market charges, on average, slightly more for a move than the move delivers, and charges even more when it knows an event is coming.

The hindsight trap

One study looked at sudden volume surges early in the trading day. On days I picked because the stock ended up big, the surge rule made +3.9% by the close. On random days, the same rule made nothing: −0.6% over 30 minutes, +0.1% by the close. The signal was the hindsight, not the volume.

The strange idea

Buying calls before news is, historically, a losing trade: you pay for a move that usually doesn't come. But Kalshi lists contracts on many of the same events: will the company say this on the earnings call, will the agency approve that. So each ticket pairs two legs.

# Ticket = $100 of the call that needs a one-expected-move rally,
#        + a Kalshi stake on the outcome that would make the call worthless.
# Size the stake so a wrong call is paid back by the hedge:
#     s · (1 − p) / p  =  100 + s      →      s = 100 · p / (1 − 2p)
# Only possible when p < 0.5, i.e. Kalshi already leans toward the bullish outcome.

stake    = 100 * p / (1 - 2 * p)
outlay   = 100 + stake
breakeven = strike + ask * (1 + stake / 100)

Every ticket lands in one of four cells: right and moved (the thesis), right but didn't move (the dead zone, where both legs lose), wrong and hedged, wrong and unhedged. The ledger reports all four, so the hedge's value is measured rather than assumed.

First results: the first four tickets to settle all landed in the dead zone. The news happened; the stocks didn't clear their strikes. Ninety-six more are open. Small sample, but it is the failure mode the design predicted, which is why the ledger tracks it by name.

Letting the rules compete

Instead of picking one version of the desk, five run side by side on paper with different settings: how big a disagreement between Kalshi and the options market triggers a ticket, how liquid the market must be, how much to hedge, when to exit. Each generation, the worst is retired and the best spawns small mutations. The first generation turned over on Oct 16.

population = [A, B, C, D, E]          # each = gap threshold, liquidity floor, hedge size, exit rule
each generation:
    score each by P&L per event cluster (correlated tickets count once)
    retire the worst
    champion  = best with enough history
    children  = mutate the best:  kinds → options-first,  max cost 12% → 8%,  gap 10 → 15 pts

What I've learned so far

• Split the data before you look. Anything chosen and scored on the same years will look better than it is. Most of my best-looking ideas died on the second half.
• Count your tries. Test enough rules and some will pass by luck. The more you sweep, the higher the bar should be.
• The boring edge is the real one. Buying dips back to the average was the only thing that held up everywhere, and it is worth about a tenth of a percent a day.
• Options charge for drama. Paying for a move before an event is paying the market's own estimate, plus a markup.
• Name the failure mode in advance. The dead zone was in the ledger before the first ticket settled, so when it showed up it was a measurement, not a surprise.

Everything here is paper trading: simulated money, no real orders. Backtests are net of the costs stated in each study; figures from the desk's files on Oct 18, 2026. This is a record of experiments, not investment advice. Questions: Discord.