Economics of the shipped strategies¶

basic_mm, avellaneda_stoikov, options_mm and xmm are reference implementations of published quoting rules, written to show the engine's API. lead_mm joins the touch of a thin pair when a liquid leader's mid, converted by an FX pair, leaves enough edge.

The example backtest and the fee table¶

configs/backtest-example.toml charges Binance spot VIP 0, [venues.sim.fees] maker_bps = 10.0:

./build/release/bin/fastmm-backtest --config configs/backtest-example.toml \
    --data synthetic --out -

backtest basic_mm  seed=1  md_events=15611  steps=16363  wall=0.01s
  net pnl                        -11.9755
  realized / unrealized / fees   0.0095 / 0.0000 / 11.9850
  fills (maker / taker)          112 (112 / 0)
  spread captured (bps)          0.008
  volume base / quote            0.19975 / 11984.99

Net PnL is realized + unrealized - fees. Trading contributed 0.0095 USDT on 11,984.99 USDT of volume: 0.0079 bps. The strategy does not read the fee table, so changing only maker_bps gives the same 112 fills and this PnL:

`maker_bps`	Fees (USDT)	Net PnL (USDT)
−0.5	−0.5992	+0.6087
0	0.0000	+0.0095
0.5	0.5992	−0.5898
1.0	1.1985	−1.1890
2.0	2.3970	−2.3875
10.0 (shipped)	11.9850	−11.9755

The break-even maker fee is 0.0079 bps. At a 0.5 bps rebate the run nets +0.6087 USDT, 98.4 % of it the rebate.

With fill_model = "matching" instead of l2_queue the same run realizes 0.0059 USDT on 8,442.57 USDT over 77 fills (0.007 bps) and gets 3 PostOnlyWouldCross rejects. The break-even fee moves; the conclusion does not.

Why the gross number is so small¶

The synthetic market has no informed flow. MarketGenerator::do_market picks the aggressor's side with a fair coin (src/sim/market_generator.cpp), and the latent mid is a separate Poisson process of ±1-tick steps in MarketGenerator::step. A market order carries no information about the next mid move, so a resting quote is never picked off, and spread_captured_bps and realised PnL are the same number: 0.008 bps × 11,984.99 USDT = 0.0095 USDT.

On a real venue the taker chooses when to trade. The quotes that fill are the ones that were wrong, so realised PnL is below the quoted half spread by the adverse-selection cost, which the synthetic market prices at zero: the backtest reports an upper bound on spread capture.

[backtest] fill_model = "l2_queue", which configs/backtest-example.toml uses, is optimistic in a second way: it estimates queue position on L2 data that has no counterparties, and it checks post-only orders against the same book the strategy saw, so it never produces the post-only rejects stale market data causes live (Configuration).

What a real session did¶

Two one-hour basic_mm sessions on BTCUSDT in Binance Demo Mode, at the account's 10 bps maker commission (Journals, replay and PnL):

Quoting	Fills	Trading PnL
15 bps from the mid	0	0.00 USDT
At the touch	1640 (all maker)	−32.94 USDT

31,622.62 USDT of notional paid 31.62 USDT in commission, and realised PnL before fees was −1.32 USDT: the fills were adverse as well as expensive.

The arithmetic you have to beat¶

A round trip pays the maker fee twice. To break even, the half spread you capture, after adverse selection, must exceed the maker fee:

captured_half_spread_bps  >  maker_fee_bps

Widening the quote does not solve it, because the fill rate falls faster than the spread rises. basic_mm's default half_spread_bps = 5 fills nothing in the synthetic market:

./build/release/bin/fastmm-backtest --config configs/backtest-example.toml \
    --data synthetic --param half_spread_bps=5 --out -

  fills (maker / taker)          0 (0 / 0)
  orders / cancels / replaces    260 / 258 / 0

configs/backtest-example.toml sets half_spread_bps = 0.01 (6 ticks at a mid of 60,000) to get fills; the Demo session above got none at 15 bps. The routes out are outside what FastMM ships: a fee tier at or below zero (maker rebate, market-maker programme, volume tier), a venue whose spread is wide relative to its fee, or a signal that makes the fills less adverse. On BTCUSDT at VIP-0 the third route is measured and closed in Judging a signal before writing a strategy.

What the backtest does not model¶

Not modelled	Consequence
Informed order flow	Adverse selection is 0
Market impact of your own quotes	Other participants do not react to you
Queue position on `l2_queue` beyond `queue_conservatism`	Fill counts are an optimistic estimate
Venue rate limits, throttles and maintenance windows	Fill and cancel counts are higher than a real session's
Funding, borrow and settlement	The backtest data carries no funding payments; perpetual and margin costs are absent (live sessions book funding on Binance USDⓈ-M and Bybit linear)
Fee tier changes with volume	`maker_bps` and `taker_bps` are constant
Commission in a third asset (for example BNB)	The engine leaves it out of fees and positions live too (Troubleshooting)

Before you size a live session¶

Run the backtest with your venue's real maker_bps and taker_bps.
Compare realised PnL against fees: if realized is smaller than fees, the strategy is paying the venue to trade.
Run the practice session in Run on a testnet or Binance Demo, then reconcile it against the account (Check PnL). A testnet's thin book will not reproduce the fill rate; Demo Mode follows the real market and charges the real commission.
Set [risk] max_loss to what you accept losing across restarts: the budget is carried in <journal_dir>/<name>.kill until you clear it (Run in production). It is not a daily budget.