Strategies

FluxrBot does not run one algorithm. It runs several, each on its own paper account with its own record, and you decide which ones you want to see. That separation is the point: strategies that share an account can never be told apart afterwards, and «the bot made money» is not an answer to «which part of it worked».

Every strategy below is described with the assumption it depends on and the way it loses. A strategy page that lists only the upside is marketing, and you would be right not to trust the rest of the product after reading one.

News → model

Reads breaking news from 37 feeds and 15 first-hand streaming accounts, asks a language model whether a story actually bears on how a specific contract resolves and in which direction, and compares that against the price in the order book. When the gap survives fees and the real book, it trades; when it does not, the reason is written down.

What it assumes: that the model reads a headline better than the market has priced it in the minutes after publication. This is the only strategy here that requires being smarter than the market rather than faster or more careful, and it is the hardest to be right about.

How it loses: the direction is wrong, or it was right and already priced. Both look identical at entry. Two defences came out of measured trades: entries are refused when the market has already moved 5¢+ toward the verdict since publication — chasing a move the crowd already made is buying its reversal — and high-confidence calls are held to resolution instead of being marked out after a day, because on settled outcomes they hit 72–94%. The system measures itself against the order-book price, not against a coin flip — beating a coin flip is not the bar.

Group arbitrage

One event, several mutually exclusive outcomes. Exactly one of them pays a dollar, so all of them together should cost a dollar. When the whole set can be bought for less, the profit is fixed at entry, whatever happens.

What it assumes: that the set of outcomes is complete — that one of the listed outcomes must happen. This is where the strategy is dangerous, and where most implementations are wrong. «Mutually exclusive» means no two can happen; it does not mean one must. Our first live run found an apparent 84¢ profit on «which state becomes the 51st» — eight outcomes trading for 15¢ together. There is no profit there: no new state may be admitted at all, and then every leg pays zero. The system now requires proof that the set is exhaustive, and treats an implausibly large edge as evidence that its own model of the payout is broken.

How it loses: a leg fails to fill and the basket is left half-built, or the set turns out not to be exhaustive. Baskets are taken whole or not at all.

Cross-venue arbitrage

The same question trades on both Kalshi and Polymarket at different prices. Buy YES where it is cheap and NO where it is expensive, and the pair pays a dollar either way.

What it assumes: that the two contracts really are the same question. Open-source bots in this space match contracts by comparing titles, and that is precisely how they lose money: Kalshi and Polymarket settle some contracts from different data sources, and «identical» contracts then resolve differently. FluxrBot has a model read the full resolution rules of both sides and answer one question — will these settle identically under every outcome — and it records the answer, including the reason for a «no». Contracts that name different settlement sources are refused outright, as are pairs whose deadlines differ, and crypto pairs are excluded entirely because the two venues use different price oracles.

How it loses: the equivalence judgment is wrong. Then the position is not a hedge but a directional bet the operator believes is safe — which is worse than an ordinary losing trade, because nobody is watching it. For that reason this is treated as a high-confidence signal, not as risk-free arbitrage, and it carries a wider margin than group arbitrage.

Forecast vs market

Daily temperature contracts are not settled by news — they are settled by a thermometer. This strategy prices them from a three-model weather ensemble (ECMWF, GFS, ICON) and trades the gap against the book.

What it assumes: that a fresh forecast ensemble knows more about tomorrow's temperature than the market price does — and that the gap between our forecast point and the exchange's settlement station stays inside the margin we reserve for it.

How it loses — and how it did: in its first week it lost a third of its paper account, and the nightly post-mortem named the cause plainly: the market was right. Every large loss had entered with a 26–47¢ disagreement against the market, which had morning observations and the actual settlement station while we had yesterday's city-centre forecast. It now refuses any trade where its own forecast disagrees with the market by more than 22¢ — a gap that size is evidence against the forecast, not against the market — and holds at most two positions at once, because a forecast error is systematic and hits every city on the same day. Those numbers are the strategy's record, and they stay on this page.

Fade the crowd

On our own settled history, YES resolves less often than its price implies — in every price band. Crowds buy the event: YES is the side of hope and of the headline. This strategy takes the boring side, buying NO in the 12–48 hours before resolution, only in price bands where that overpricing has been measured on at least thirty settled outcomes, and only where it survives fees.

What it assumes: that the measured base rate persists. It needs no model and no news — the edge is not speed or reading, it is the willingness to hold the dull side. The overpricing is re-measured on every run from our own resolved contracts: if the market corrects, the strategy goes quiet on its own, and a band without enough history is not traded at all.

How it loses: adverse selection — the one contract where the price is held up not by the crowd but by somebody who knows something. Thin books are refused for exactly that reason. And NO positions are correlated: a wave of surprises hits all of them on the same day, which is why it holds at most five at once.

What we do not run

Market making, wallet copying, and speed trading on live sports are not implemented, and the reason is the same in each case: none of them can be tested honestly on paper. Market making needs live resting orders, wallet copying is a separate on-chain problem, and sports speed needs paid real-time data. When something cannot be verified before it touches money, it does not ship.

What this page is not

None of this is a promise of profit. Every strategy runs on a virtual $100 account first, and the record it accumulates — including the trades it declined and why — is what decides whether it is worth anything. If a strategy does not earn its place, it is switched off and that is written down too.