Comeback Cicada

About

Comeback Cicada is the only agent here that acts on a game already under way. It watches a home team losing early and asks one narrow question: is the live price lower than the share of real games that reached the same state and ended with the home team winning?

The comparison is against a count, not a forecast. Bottom of the first, nobody out, bases empty, home team down three: 1,998 of 6,857 games in the Retrosheet play-by-play record ended with the home team winning, which is 29.1 percent. A market quoting that state at 19 cents is either reading something the count cannot see, or it is wrong. This agent is the measurement, not a claim that early comebacks are underpriced.

It is worked by hand. The exchange's market feed carries prices and nothing about the game itself, so a person reads the inning, the outs, the runners and the score, checks the table, and enters the pick under the same published rule.

The rule

Watch a live game where the home team trails. In the bottom of the first, second or third inning, read the home team's Kalshi ask and compare it with the Retrosheet base rate for that exact state: deficit, inning, outs and runners. If the ask sits at least 8 points below that base rate and is 25 cents or less, buy the home team at the ask. Sell at the entry plus 7 cents. If the sell does not fill, hold to settlement. If the gap is under 8 points, no pick.

Worked by hand. This rule reads the inning and the score, so a person reads them from the live game and enters the pick exactly as the rule says. Every pick is graded the same way as the rest of the Lab.

How it works

• Operated by hand. The registry marks it manual, so the scheduled runner skips it before evaluating anything: the markets API carries quotes, not inning, outs, runners or score.
• The base rates are counts, not a model. The table is a verbatim subset of Greg Stoll's probs.txt, the data file behind his Win Expectancy Finder, generated from Retrosheet play-by-play. Its sha256, its read date and its span of 1903 to 2025 across 195,573 games are recorded in the source. The information was obtained free of charge from and is copyrighted by Retrosheet.
• Only states the rule can reach are embedded: home team batting, innings one through three, deficits of one to nine runs. A rate is returned only at 300 games or more. Nothing is interpolated, smoothed or filled in from a neighbouring cell.
• The state has to be complete and exact: bottom of the inning, inning 1, 2 or 3, zero to two outs, a readable runner mask, integer scores on both sides, and a home deficit above zero. A partial feed is a refusal, because a missing outs count read as zero would be an invented game.
• It identifies the home market by elimination rather than by name, because the rules text cuts multi-word home names short.
• Then the arithmetic: the ask must be 25 cents or less, clear the same 100 resting and 3 cent spread gate, and sit at least 8 points under the base rate. The row carries the state read, the base rate compared against, and a sell target of entry plus 7 cents.

It will not estimate a base rate it does not hold: a state outside the table, a cell counted off fewer than 300 games, or a feed that answers only partly is a refusal rather than a number, because one invented base rate would make every comparison after it meaningless.

What would prove it wrong

If the trailing home teams it buys on an 8 point gap go on to win at about the share the market priced rather than the share the Retrosheet count implies, the early price was not overreacting and the premise is wrong.

Why this rule

Tests whether a live in-game price overreacts to an early deficit. The base rate for a home team down three in the bottom of the first is 29.1 percent across 6,857 Retrosheet games; a market that prices it at 19 cents is either seeing something the count cannot, or it is wrong. This agent is the measurement. It is NOT a claim that comebacks are underpriced. Operated by hand for now: the rule needs live inning, outs and deficit, which the Kalshi markets API does not carry, so the runner cannot fire it without a score feed.