How Arena fills, grades and ranks

Market orders

A market order fills at the live price at the moment Arena’s server receives it: a buy at the live ask for your side, a sell at the live bid. The server reads the price itself. A client can send a ceiling on a buy or a floor on a sell, which refuses a worse fill, but never the price to fill at.

The live price is Kalshi’s, unless the trade asks for a best-price fill and another venue passes Arena’s checks for the same instrument (see Best-price fills below). Either way the receipt names the venue.

There is no contract limit: your balance is the only limit on size. The one exception is what Arena can store: one trade records at most 1,000,000,000 contracts, and a trade is refused when your balance plus everything your open trades, resting buys and this trade could pay would pass 9,000,000,000 paper dollars. The refusal reads “That trade is larger than Arena can record.”

There is no depth walk, no queue and no partial fill: the order fills whole at one price, or not at all. Paper trades pay no fee. The eval board’s grades subtract the fee Kalshi would charge (see The eval board). The stake is contracts × price ÷ 100, and a sale pays contracts × bid ÷ 100.

A buy is refused, and nothing is charged, when your spendable balance cannot cover the stake, when the market is closed or already resolved, when no one is offering your side, or when the ask is above your ceiling (--max-price in the CLI, max_price_cents on the MCP server).

Opening a position is for Arena Basic members. A 402 (CLI exit 5) means the account needs a membership. Selling is not limited to members.

A sell always closes the whole position. It is refused when there is no bid to sell into, when the market has already resolved (the trade will settle instead), when the position is a leg of a parlay or list, and when the position was opened before September 2026 and carried (see Known limits).

Limit orders

A limit order names your price, in cents of the side you trade. If the live price already meets it when the order arrives, it fills at once, at the live price. Otherwise it rests.

A resting buy holds contracts × limit ÷ 100 out of your spendable balance, and a limit buy your balance cannot cover is refused. A resting sell holds nothing, because the position is the collateral, and only one sell order can rest on a position.

Arena checks resting orders against the live price about every 30 seconds. A buy fills when the live ask for its side is at or below the limit, a sell when the live bid is at or above it. The fill is at the live price, so it can be better than the limit. On a buy the hold comes back and the actual stake is taken, so a fill can cost less than the hold.

An order fills whole or not at all. There are no partial fills.

When a limit order asks for a best-price fill and can fill the moment it arrives, it is checked like a market order and can take another venue’s price. An order that rests fills later at Kalshi’s price only.

• gtc, the default, rests until it fills, expires, is cancelled, or 365 days pass.
• ioc and fok never rest: if the price is not there when the order arrives, the order is cancelled at once. With no partial fills, the two behave the same.
• --expires in the CLI (expires_at on the MCP server) sets when a resting order stops, at most 365 days out.
• A resting order is also cancelled, with any hold returned, when its market closes or resolves, and a resting sell when its position closes first. While the live price of a market cannot be read, its orders are neither filled nor cancelled because of price, though expiry still applies.
• arena cancel (cancel_order) withdraws a resting order and returns any hold in full. Cancelling an order that already filled is not an error: the answer names the trade it became.

Your probability

When you place a trade you can state your probability, in whole percent from 1 to 99, that the side you are buying wins. A NO buy states the chance that NO wins. It is a statement, not a price: it never changes the fill, the stake, the payout or any guard.

Where to give it: the “Your probability” field on the website’s order ticket (not on a Combine account), --prob in the CLI (arena buy <ticker> --side yes --contracts 10 --prob 62), and stated_prob_pct on the MCP server’s place_trade. The iPhone app adds it in a later release. A limit order keeps it while it rests, and its fill carries it to the position. Tickets and receipts show it against the price as an edge: 62% against a 55% price is +7% on the website (the CLI and SDK print +7 pts). If the server could not save it, the ticket and the receipt say so; the trade stands.

How it is scored: once a pick settles and the eval board shows it, the board scores your stated probability (Brier score and log loss) and labels the pick stated. An account whose every scored pick in a window has one is ranked on log loss; otherwise on closing-line value or edge, as before. The AI models now state a probability on every pick; their earlier picks are scored on their confidence and labelled confidence, so a window that spans the change reads mixed for a model.

Who sees it: you, on your own positions and history. On the eval board it appears only for settled public picks, by the same rule as the pick itself. Open picks never show it.

Drawdown in eval rows

Eval rows (the public board’s get_eval_board and your own row from get_my_eval) state drawdown two ways: the worst fall of realized P&L in paper dollars, and the worst fall of a time-weighted index of paper equity in percent. The two can come from different falls. In the Python SDK 0.2.0, eval_board() rows carry the two only in each row’s raw dict, for example row.raw["max_drawdown_dollars"]. Python SDK 0.3.0 names them: row.max_drawdown_dollars and row.max_drawdown_percent.

• max_drawdown_dollars: the largest peak-to-trough fall of realized P&L, in paper dollars, on the same basis as pnl_points. Picks count in settlement order (closed_at, then the trade id), the running total starts at 0, and the peak never sits below 0, so you can rebuild it from the account’s own pick rows (pnl_points, closed_at, trade_id).
• max_drawdown_percent: the largest fall in a time-weighted index of your paper equity, to three decimals. A reset’s refill counts as a deposit, not a gain, so it stays between 0 and 100. It is a percent, not a fraction: 0.198 means 0.198%, a $198 loss on a $100,000 start.
• How the index works: each settled pick is one step, and its return is its P&L on the equity the curve had just before it. The index starts at 1 and multiplies those returns, and the percent is its largest fall from its own highest point. Equity is the balance the curve stands on plus the realized P&L of the row’s picks since then.
• It is an index on the row’s picks, not the account’s real balance. A 7d, 30d or 90d row starts from the grant ($1,000 before the Sep 27, 2026 bankroll change, $100,000 after it, $100 for a month before September), not from what the account held when the window opened. Only the row’s picks move it: public picks on the board, plus your private picks on your own row; parlays are not in it.
• Two things count as deposits, not gains: the +$99,000 top-up of the Sep 27 change (it reached balances on Sep 28 at 08:34 UTC) and a reset’s refill to $100,000. A deposit changes the balance the next returns are measured on, never the index, so a fall that runs across one is one fall. A pick placed before a deposit and settled after it counts on the new balance.
• A reset never hides a drawdown. A loss after it still deepens the fall from the old high, and a gain after it only sets a new high once the index passes the old one: an account that loses 90%, resets, then loses 90% again reads 99. The percent reads 100 once the equity on the row’s curve reaches $0, and it stays 100 after that. That is the curve’s equity, not the account’s: a 7d, 30d or 90d row leaves out profit from before the window, and a stake can be funded by P&L that is not on the row, such as a parlay win, so a row can reach $0 while the account holds more. It is not a bust on the leaderboards (a run that keeps under 10% of its start at a reset): a run that loses 95% and resets reads 95.
• Order: picks count in settlement order (closed_at, then the trade id), one at a time, the same order as the dollars, so picks that settle at the same moment are separate steps in trade id order. A pick that settles at the moment of a reset belongs to the run the reset ends: the positions a reset writes off are stamped with its time and are losses of that run.
• The board row others see has the percent only while the account’s balance numbers are shown (a public account with no hidden trades or resting buys). Otherwise its max_drawdown_percent is null, whether or not the account ever reset, so it never reveals a hidden reset. Your own row from get_my_eval always has it and counts every one of your resets. While your numbers are shown, your board row and your own row carry the same two numbers.
• max_drawdown_points and max_drawdown_pct are deprecated. They are on a scaled curve on which a pick placed before the Sep 27 change counts at 100 times its P&L (it was sized on the $1,000 grant), and max_drawdown_pct is a fraction (0 to 1) of a peak on that curve. They keep their values because a signed /promote record carries the same two numbers, and they will be renamed with the next record version. Use the two fields above.

Since Sep 1, 2026 each account has one curve that neither a month nor a reset restarts. Before that, each calendar month (New York time) was its own curve, because balances restarted every month. Each field is the largest of its kind across the curves, so the dollars and the percent can come from different falls. max_drawdown_stakes (the same fall in the account’s own median stake) and drawdown_segment (ledger, or a pre-September month such as 2026-08) are unchanged.

Best-price fills

A paper trade can fill at Novig’s, Polymarket’s or ProphetX’s price for the same instrument instead of Kalshi’s. It is still a sandbox fill: it is recorded in Arena’s ledger and not sent to the venue.

A trade asks for it. On the website, the market page’s order ticket lists the venues and opens on the cheapest one for the side (every venue at that price when several tie); the trader can tick more or untick any, and Kalshi alone means Kalshi only. The terminal’s quick ticket has a “Best price” or “Kalshi only” choice, and Best price is its default. In the CLI it is --venue best, and on the MCP server route: "best", both the default from 0.5.0. The iPhone app sends no route, so its trades fill at Kalshi’s price until an app update. --venue kalshi and route: "kalshi" always fill at Kalshi’s price, for a bot that models Kalshi’s book alone.

Best-price fills and their preview are for Arena Basic members. A free first pick fills at Kalshi’s price, and so does a sell from an account without a membership; the answer says why.

Before you confirm, the website and the CLI show where the fill would be priced now, with Kalshi’s price beside it and the reason. On the MCP server, dry_run: true answers the same and places nothing. A preview is not a promise: the server decides again when the trade is placed, and the receipt names the venue that filled it.

Another venue’s price wins a buy only when all of these hold. Otherwise the trade fills at Kalshi’s price, and the reason is recorded:

• The venue’s match meets Arena’s rule for showing a match next to Kalshi: matched by rule with a confidence of at least 0.9, not delisted, and either exact or differing only in a named case (see /docs/methodology). Novig’s NHL moneylines and totals are left out until Arena has read Novig’s NHL overtime terms.
• The venue’s price is at most 15 seconds old at the source.
• It is within 5 points of Kalshi’s midpoint, and Kalshi shows both a bid and an ask. A price that far from Kalshi’s is more often a stale quote or a wrong match than a price anyone could get.
• It is at least 1 point below Kalshi’s ask, after the venue’s ask is rounded up to a whole cent.
• It is above Kalshi’s bid. A price at or below Kalshi’s bid usually means one book is stale, so it is refused, never moved to Kalshi’s bid.
• The venue shows at least the number of contracts you asked for at that price, less what Arena has already filled against that same quote in the last few seconds. A venue that shows no size there does not win.
• The game has not started by its scheduled start time. Best-price fills are pregame only.
• The venue answers in time (about 1 second each, 1.5 seconds in all), its switch is on, Arena’s own pace for reading that venue is not spent, and the account has asked for no more than about 20 best-price checks in the last minute (about 60 from one IP address across accounts).

So another venue wins only strictly inside Kalshi’s own spread, at the price that venue showed, and on a 1-point Kalshi spread no other venue can win. A best-price fill beats Kalshi’s ask by less than Kalshi’s spread, and a buy followed by a sale at Kalshi’s bid cannot make money. The lowest price that passes wins. A tie goes to a match a person has reviewed, then to Novig before Polymarket.

A sell mirrors this on bids: the highest bid that passes, at least 1 point above Kalshi’s bid and below Kalshi’s ask. A position filled on Kalshi exits on Kalshi, because no Novig or Polymarket match is exact today. A position filled on Novig or Polymarket can exit on that venue or on Kalshi, and every exit says how its contract differs.

Parlay legs, trades placed into a list, Combine evaluation accounts and resting limit orders always fill at Kalshi’s price. Fills stay whole size, with no fee and no depth walk, whatever the venue.

Every fill placed through the website, the CLI or the MCP server records its venue, except on a Combine evaluation account, which always fills at Kalshi. A trade that asked for a best-price fill also records the venue’s market and side, Kalshi’s price at that moment and the reason, whichever venue won. Receipts, positions and history show the venue. A trade with no venue recorded was priced at Kalshi: every trade from before best-price fills, Combine evaluation trades and the AI models’ picks.

Arena can switch best-price fills off at once, for every venue or for one. Trades then fill at Kalshi’s price.

Settlement and grading

A trade settles from its market’s result. A YES contract pays $1 if the market resolves yes and nothing if it resolves no, so a winning trade is paid contracts × $1 and a losing trade loses its stake.

A market that ends with no yes or no result, for example because it was cancelled, is graded void, like a push: the stake comes back and the P&L is zero. A market that resolves to anything other than yes or no is held for a manual review instead of being graded automatically.

A trade that took another venue’s price settles from the Kalshi market’s result too, exactly as a trade filled on Kalshi does, even where the two contracts differ. A Novig NFL moneyline voids when the game ties and Novig settles the void by its own rule, while Kalshi settles a tie at 50. On Arena a tie settles every trade on that market at 50, whichever venue priced it, and waits for the review above. The same holds for every other known difference (a game that is not completed on Novig, a postponed or cancelled game on Polymarket, Polymarket NFL terms not yet compared). The receipt’s rule note says where the contracts differ.

Every settled trade is won, lost or void. A trade closed early by selling is sold: it has a P&L but no grade.

Arena looks for results every five minutes, and sooner when it hears that a market has been decided, so a trade usually settles within about five minutes of the result. arena settle (settle_open on the MCP server) settles your own resolved trades at once, as a backstop.

Net worth

Arena has one definition. Net worth is your spendable balance, plus what you paid for your open positions, plus the holds of your resting buy orders. Parlay legs and Combine evaluation trades are not part of it.

Open positions count at cost, not at the live price, so opening a position or resting an order does not change net worth. Only a sale or a settlement does. The leaderboards rank profit: all time across every run, or from trades closed in the last 30 or 90 days. See Balance and resets.

Balance and resets

Every account starts with $100,000 in paper dollars. Your balance carries over. Resting orders stay until they fill, expire or are cancelled.

You can reset your account at most once every 30 days. Turn resets on in Settings first (saved accounts only, not guest sessions). A reset sets your cash back to $100,000 and starts a new run. Resting orders are cancelled. If you have open positions, sell them first or choose to forfeit the ones you see. A forfeited position counts as a loss, and if your open positions change before the reset goes through, it stops and asks you to check them again.

Resets show on a public profile. Your profile shows each reset and the result of each run. Leaderboards show how many times you reset and how many runs went bust. A run goes bust when what it keeps at the reset, cash plus the money in resting orders, is below 10% of its start. A forfeited position is lost, so it does not count.

Leaderboards. All time ranks total profit since September 1, 2026, across every run. It lists accounts with at least 20 closed trades and a first trade 30 or more days ago. 30 days and 90 days rank the profit of trades that closed in that window, parlay legs and forfeits included, for accounts with at least 5 closed single trades in it. A forfeited position counts as a closed trade and a loss. Open positions count at cost, so opening one never changes your profit. Your rank also moves as trades age out of a window and as other traders close theirs. The boards update once a day, from the daily balance snapshot at about 05:10 UTC, so a closed trade shows on them after the next snapshot.

Use one account per model. A reset is a fresh start for the balance, not a new model version: the 30-day and 90-day windows and all-time profit still include the trades from before it.

What “recorded by Arena” means

• Server-priced: the price comes from the live market at the moment Arena’s server fills the order, never from the client, and the trade records which venue’s price it took.
• Server-timestamped: the database stamps when each trade opens and when it closes.
• Graded from the market result: a server job settles each trade from its market’s result. The trader never chooses the outcome.
• Not editable: no client can set or change a trade’s price, its open and close times or its result, or delete a single trade. Only Arena’s server functions write those. You can still file, share and group your own trades, and deleting your account removes them all.

It does not mean a third party has audited the record. It means Arena’s own engine priced, timed and graded it by the rules on this page.

Estimated venue fee

Paper trades pay no fee. The eval board’s grades subtract the fee Kalshi would charge (see The eval board). So that a bot can see what the same order would cost on the real venue, quotes can show venue_fee_est_dollars: an estimate of Kalshi’s taker fee from its published formula, labelled “Estimated venue taker fee. Not charged on paper.” From CLI 0.6.0 the CLI’s note is the plain one: paper trades pay no fee, and the eval board’s grades subtract the fee Kalshi would charge. arena market <ref> --contracts N, the MCP tool get_market with contracts and an MCP place_trade dry run that would fill on Kalshi show it; an MCP limit dry run shows it only when the limit would fill now on Kalshi, since a resting order is not a taker, and one that would fill on another venue leaves it out. From CLI 0.6.0 the CLI’s arena buy … --dry-run does not show it: that dry run is the server’s preview of the paper fill and adds no estimate of its own. It never changes a paper trade. From CLI and MCP server 0.6.0 it uses Kalshi’s series multiplier too (0.5 on MLB games, spreads and totals when Arena read them on 2026-09-27), as the eval board does; 0.5.0 uses Kalshi’s general rate, so on an MLB game it shows about twice the fee /docs/methodology works out. Each venue’s fee formula, with the multiplier, is on /docs/methodology.

An MCP place_trade dry run that would fill on Kalshi, and get_orderbook with contracts, can also show what buying that size through Kalshi’s book would cost (walk_vwap_cents and walk_slippage_cents on the dry run, walk on get_orderbook). From CLI 0.6.0 the CLI’s dry run does not show it. Arena itself fills the whole size at one price.

Known limits

• Fills ignore depth. The whole size fills at the top price, where a real order that size could move the price.
• Paper trades pay no fee, so paper P&L leaves out what a real account would pay in fees. The eval board’s grades subtract the fee Kalshi would charge: the eval board shows P&L after fees.
• A best-price fill takes the price another venue showed Arena at that moment. A live order there could get a different price or none, and some people cannot use that venue at all (see /docs/methodology). Best-price fills are pregame only, and their checks look at each venue’s top price, not deeper in the book.
• Limit orders fill on touch, as if they were first in the queue, and are checked about every 30 seconds, so a price that touches the limit and moves away between two checks can be missed.
• No partial fills, and a sell always closes the whole position.
• Markets priced in fractions of a cent are refused, and orders resting on them are cancelled.
• A position opened before September 1, 2026 that was still open that day keeps its result on your record, but its payout or refund is not added to your balance. It cannot be sold either: a sell on it is refused, and a limit sell on it is cancelled when it would fill. It settles on its own. Positions opened since then settle normally.
• Open positions count at cost on the leaderboards, so an open position neither lifts nor lowers a rank until it settles or you sell. A position you forfeit at a reset counts as a loss of its whole stake.
• A market that resolves to something other than yes or no waits for a manual review before its trades settle.