How Rivo measures whales, traders and strategies.

Two lanes, one history

Every trade Rivo records lands in one of two lanes, and some in both. The whale tape holds trades above a size floor, whoever placed them. The traders feed holds any trade, of any size, by an identified trader: a Polymarket wallet or a Kalshi nickname. The tape answers “where is big money moving”; the traders feed answers “what did this account just do”. Both write to the same durable history, and everything downstream reads from it.

Position classification runs on Polymarket trades before they are recorded. Each trade is compared with the wallet's holdings in that market:

Settlement and the win rule

A resolver polls each venue twice a day for every market with an open recorded position. When a market resolves, every recorded buy in it is scored: buy_yes wins on a yes resolution, buy_no on a no. A market that is still technically open but trading at 99.75 percent or higher on one side is treated as settled for that side, because the remaining tail is exchange mechanics rather than uncertainty. Markets that are voided or cancelled are marked void and excluded from every rate and every PnL figure.

Sells are never scored. A sell reduces a position that may have been opened before Rivo saw the wallet, at a price Rivo did not record, so any PnL assigned to it would be invented. Sells are recorded and drive followed-trader alerts and Autopilot exits, but they carry no resolution. This is why every win rate on the site is a win rate on buys.

Paper PnL

For a recorded buy at price p with a dollar stake S, the position holds S / p contracts. On a win each pays one dollar, so profit is S / p − S. On a loss the stake is gone, so profit is −S. On trader profiles and leaderboards S is the trader's actual fill size; in backtests and research it is a fixed stake, $100 by default, so two records can be compared without the larger bankroll winning by default.

The last column is the one most people skip. A 70 percent win rate is excellent at 40 cents and a losing record at 80 cents. That is why every Rivo surface shows the entry price next to the outcome, and why the strategy builder lets you bound a rule by price band rather than by win rate alone.

Grading traders

Each identified trader carries a copy score from 0 to 100 and a grade from A to D. The score is the lower bound of a 95 percent Wilson confidence interval on the trader's win rate, computed over resolved recorded trades from the last 90 days, multiplied by 100. Fewer than five resolved trades in the window means no score and no grade. Grades are fixed cut-offs: A at 60 or above, B at 50, C at 40, D below that.

The Wilson bound was chosen because it folds sample size and accuracy into one number without a hand-tuned weight. A perfect five-for-five record scores lower than thirty wins from forty, which matches how a careful human would read those two histories.

The profit figure on a profile is the first available of: observed net paper PnL on trades Rivo recorded, the venue's official account PnL where one is published, or the current value of open positions. The profile labels which one it is showing. Observed PnL is computed at the trader's real fill sizes, so a large bankroll shows large dollar figures; the flat-stake replay on the same page is the number to use when deciding whether to copy them at your own size.

Backtesting a strategy

A strategy is a named set of criteria over recorded trades: venue, category, trade-size range, entry-price band, direction, position type, market keywords, and sports league. The same matching function runs the live strategy matcher and the builder's preview, so a rule that matches in the preview is the rule that will be recorded. Strategy matching only considers opens, adds and flips; trims and closes never create a strategy event.

A backtest replays that rule over a window, by default the last 90 days, and reports:

• matched trades, and how many of them have settled;
• wins, losses, and the win rate on settled trades;
• total paper PnL at the chosen stake and the return on stake, which is PnL divided by stake deployed on settled trades;
• trades still open at the end of the window, counted separately and never scored;
• with holdout on, all of the above twice: once for the earlier two-thirds of the window and once for the later third.

The stake resolves in a fixed order: a per-trade override, then the strategy's default stake, then the whale's own size. A stake of zero on a single trade is respected as “I would have skipped this” and does not fall back to the whale's size. Stakes are deliberately uncapped; the interface flags a stake larger than the whale's own fill rather than forbidding it.

Holdout, sweeps and breakdowns

A rule tuned until it looks good on all the data is fitted to that data. The holdout scores the rule on the earlier two-thirds of the window and validates it on the final third, which the tuning never saw. The rule holds when the out-of-sample return is above zero. A rule that is strongly positive in sample and negative out of sample is a fitted rule, however good the headline looks. This is a guard against the most common mistake, not a significance test, and re-tuning after peeking at the holdout defeats it.

A sweep scores one rule across up to twenty values of a single numeric field in one call, which shows whether a result depends on a narrow threshold. A breakdown splits one rule's results by category, venue, month or price band. Both are charged per variant against the API quota, and both are available through the MCP server as well as the terminal.

What a paper result leaves out

A paper result assumes a fill at the recorded price for the full stake, held to settlement. That makes it reproducible; it does not make it executable. The gaps, in rough order of how much they usually matter:

• Entry delay. By the time you see a trade the price has often moved. A $100 buy at 40 cents returns $150 on a win; the same buy five cents later at 45 cents returns $122. The outcome did not change, the payoff fell by a fifth.
• Spread and depth. The recorded price is the whale's fill. Whether the book has that size at that price a moment later is a separate question, which is why connected orders in the terminal quote first.
• Fees. Both venues charge fees on fills. Paper PnL is gross.
• Correlation. Ten recorded buys in one market are ten events but one outcome. Distinct-market counts are shown everywhere for this reason.
• Capital. A flat stake on every match ignores how much would be tied up at once. Autopilot's per-day cap exists because real accounts are finite.
• Survivorship. Choosing a trader because they already rank highly is a selection. The holdout and the 90-day scoring window limit how much a single hot streak can dominate.

None of this makes paper results useless. It makes them a filter: a rule that cannot clear a gross paper test will not clear a net live one, and a rule that does still has to earn its place through a paper strategy before Autopilot spends a dollar on it.

Published research

Studies on the research page apply these same definitions to a fixed observation window and freeze the result. Each states its population filter, window boundaries in UTC, stake, method version and export time, and ships the aggregate table as JSON and CSV with a SHA-256 content hash. A study is never edited after publication; a re-run of the same window at a later date is a new study with a new date, because settlements continue to arrive.

Glossary

• Whale tape: trades above the size floor, any trader.
• Traders feed: any trade by an identified trader.
• Position delta: open, add, flip, trim, close or unknown.
• Resolution: win, loss, void, or unresolved.
• Paper PnL: stake / price − stake on a win, −stake on a loss.
• Return on stake: paper PnL divided by stake deployed on settled trades.
• Copy score: Wilson lower bound of the 90-day win rate, 0 to 100.
• Holdout: earlier two-thirds to score, final third to validate.
• Strategy event: one recorded match between a trade and a strategy.

Frequently asked questions