Market making

Post a bid a little below the mid and an aska little above it. If a seller hits your bid and, later, a buyer lifts your ask, you've bought low and sold high — pocketing the spread without ever predicting the price. Do this thousands of times and the edge compounds. On exchanges with a maker rebate(you get paid for providing liquidity), you earn that on top. The Arena's real-book season uses a maker rebate and a taker fee, so providing liquidity is rewarded and crossing the spread costs you — exactly like a real venue.

• Inventory risk. Every fill leaves you holding a position. If your bid keeps getting hit, you accumulate a long inventory right as the price may be falling — your spread profits get swamped by a directional loss. You must lean against inventory.
• Adverse selection.Your quote gets filled precisely when someone knows something you don't — the market is about to move through your price. Makers are systematically picked off by informed flow; the spread you earn has to compensate for it.
• Queue position. At a given price, fills go first-in-first-out. Re-quoting every tick is simple but sends you to the back of the line; quoting patiently keeps your place. (See how matching works.)

Here's the core of a working market maker (full file on GitHub). It cancels its old quotes, recomputes a bid and ask around the mid, and skews both quotes against its current inventory so fills naturally pull it back toward flat:

from convexpi.arena.client import RemoteAgent, MarketState

class MarketMaker(RemoteAgent):
    half_spread_bps = 8.0   # how far each quote sits from the mid
    size = 5                # quote size per side
    max_pos = 40            # hard inventory cap
    max_skew_bps = 8.0      # max quote shift at full inventory

    def on_tick(self, state: MarketState) -> list[dict]:
        if state.mid is None:
            return []
        orders = [self.cancel(o["order_id"]) for o in state.my_open_orders]

        mid  = state.mid
        half = mid * self.half_spread_bps / 1e4
        # lean against inventory: long -> shift quotes down (sell more, buy less)
        skew = (state.position / self.max_pos) * (mid * self.max_skew_bps / 1e4)

        if state.position < self.max_pos:
            orders.append(self.limit("buy",  round(mid - half - skew), self.size))
        if state.position > -self.max_pos:
            orders.append(self.limit("sell", round(mid + half - skew), self.size))
        return orders

That's the whole strategy. When flat, the quotes are symmetric around the mid; as inventory grows, the skew shifts both quotes to unload it, and the position cap stops either side from running away.

Install the SDK, then point the agent at the real-order-book competition:

pip install convexpi-arena

python examples/market_maker.py my-handle \
    --server wss://arena-production-e3f1.up.railway.app

Watch your fills print, and your PnL and maker % climb the board on /compete/arena-book— where you can also see the live depth ladder you're quoting into.

The starter is deliberately naive. Ideas that move the needle:

• • Keep queue priority: only re-quote when the mid moves enough to matter, instead of cancelling every tick.
• • Widen in fast markets: scale your half-spread with recent volatility so you're not picked off during moves.
• • Smarter skew: make the inventory penalty non-linear, or quote different sizes on each side.
• • Respect the fees: your spread must clear the round-trip cost after the maker rebate — quote too tight and you pay to trade.
• • Read the tape: back off when recent trades show one-sided, informed flow.