How the AI Traders Decide

Learning objectives

After reading this page you will be able to:

• Follow one AI trader from a decision to the orders it sends
• Explain each strategy, execution style, tempo and risk setting
• Predict how a preset will behave, and write your own

Prerequisites: AI Traders.

The loop

An AI trader does nothing on a fixed timer. Decisions arrive at random — a Poisson process at the tempo's decisions_per_min — so a swarm's orders do not come in bursts. In between, it keeps its house in order. Each time it wakes, in this order:

1. Housekeeping. Cancel resting orders older than max_order_age_sec, or more than stale_price_ticks behind the touch on their own side (a buy order behind the best bid, a sell order behind the best ask).
2. Slices due of a twap parent order.
3. Protection. Place, resize or remove the protective orders for each open position (continuous trading only).
4. A decision, if one is due: the strategy names a symbol, a side, an urgency and a size multiplier — an intent — or nothing. The execution style turns the intent into orders, within the risk limits.

During continuous trading the decision rate follows the day's U shape: about twice the average after the open and before the close, half of it at midday. A swarm's budget scales every trader's rate so the swarm as a whole sends the actions per second it was given.

The trader reads the market from the engine's public feed — book snapshots, trades, the session, halts — and from pm-market-sim the true values and the news. It sends nothing for a halted symbol, nothing while the market is closed, and only what each phase allows: resting LIMIT and ICEBERG orders in the call phases, ATO in the opening auction, ATC in the closing auction, everything during continuous trading.

Strategies: what to trade

A strategy says which way it leans with a score from −1 to 1 and buys with probability (1 + score × strength) / 2, so it leans without becoming predictable.

Strategy	Looks at	Buys when	Parameters
noise	one random symbol	at random (score 0)	urgency
trend	sample random symbols, acts on the strongest signal	its fast average price (30 s) is above its slow one (300 s)	strength, sample, urgency
reversion	the same way	the price is below its slow average	strength, sample, urgency
value	sample random symbols, acts on the widest gap	the price is below its view of the true value by more than threshold	threshold, bias_std, redraw, sample, urgency, rumour_shift
news	the news touching its symbols	the headline is good news	lag_sec, lag_sigma, half_life_sec, threshold, urgency, move

Trend and reversion need about 30 seconds and 20 prices of a symbol before they have an opinion on it. Both scale the difference by the symbol's own recent volatility, so a 1% move means more in a quiet stock than in a wild one. The stronger the signal, the more urgent the intent.

Value sees each symbol's true value through a bias of its own — a persistent error, drawn per symbol with standard deviation bias_std and re-drawn now and then (redraw). That is why two value traders disagree, and trade with each other. It acts once the log gap between its view and the price exceeds threshold; the wider the gap, the more urgent and the larger the order, up to three times its usual size. When it crosses the spread it pays up to its view of the value less threshold — not one tick past the touch — so a jump in the value is priced in by a few orders rather than a thousand one-tick steps. An open rumour shifts its view by rumour_shift × sentiment × credibility.

News hears each headline after a delay of its own (lognormal, median lag_sec), so news traders do not all hit the book in the same millisecond. From then on its conviction is the headline's |sentiment| — times the credibility for a rumour — halving every half_life_sec. While conviction is above threshold it trades the headline's symbols in the headline's direction, urgently and in up to three times its usual size, paying up to move × conviction (a log move) beyond the price it saw when it first acted on the headline. A confirmation turns a rumour into news known to be true; a retraction leaves an impulse the other way, as large as the rumour was believable.

Execution styles: how to trade it

Style	Sends	Parameters
passive	A LIMIT offset_ticks behind the touch (0 joins it)	offset_ticks, tif
marketable	A LIMIT cross_ticks through the touch; what does not fill rests	cross_ticks, tif
sweep	A MARKET, IOC or FOK order against the other side	sweep_type, cross_ticks
iceberg	An ICEBERG behind the touch showing visible_fraction of its size	offset_ticks, visible_fraction
twap	A parent order cut into slices child orders over horizon_sec	slices, horizon_sec, child_style

Any style becomes marketable when the intent's urgency reaches urgency_cross. A crossing order (marketable, IOC, FOK) is priced at the strategy's limit when it gives one (value and news do), otherwise cross_ticks past the touch. In the call phases — pre-open and the auctions — a trader takes part in a decision with probability auction_participation, and then sends a limit order straddling the reference price (with TIF ATO or ATC in the auctions), so that buyers and sellers overlap at the uncross. Prices are kept inside the symbol's price collar and quantities inside its order limits: an AI trader does not send what the engine would refuse.

Tempo: how often and how big

Tempo	Decisions per minute	Order size	Sizes
many-small	30	1–25	mostly small
aggressive	20	20–120	even
cautious	6	10–60	even
few-large	2	150–700	mostly large

A preset can give its own instead: tempo: {decisions_per_min: 12, size_min: 5, size_max: 80, size_distribution: small-heavy} (balanced, small-heavy or block-heavy). The strategy's size multiplier applies on top.

Risk: what it will not do

Setting	Effect
max_position	No order that could take the position past ± this many shares, counting open orders as if filled
max_live_orders_per_symbol	At most this many working orders per symbol
max_order_age_sec	Cancel resting orders older than this
stale_price_ticks	Cancel resting orders this far behind their own side's touch
protection, protection_pct	Protective orders for open positions, protection_pct of the price away

protection is one of none, stop (a STOP order), stop_limit, trailing (a TRAILING_STOP) or bracket (an OCO of a take-profit LIMIT and a STOP). It follows the position: a new fill resizes it, a flat position removes it.

A trader that gets 25 orders refused within 10 seconds pauses for 5 seconds — the reject breaker — so a misconfigured preset cannot flood the exchange with refused orders. A cancel that lost the race with a fill, or an order in flight when the market closed, does not count.

Writing a preset

Start from the closest built-in preset in src/edumatcher/ai_trader/presets/ and change one part at a time:

You want	Change
More flow	tempo
Takes liquidity instead of providing it	execution.style: sweep or marketable
Holds bigger positions	risk.max_position
Reacts to news faster	strategy.lag_sec (news), execution.urgency_cross
Disagrees more with other value traders	strategy.bias_std (value)

Every parameter has a range; a value outside it is refused with the file and the key named. AI Traders shows how to run your preset.

The other bot: pm-mm-bot

pm-ai-trader trades like a directional participant. The market-maker bot has a different job: it posts two-sided quotes on many symbols from one participant ID. Run one with the swarm — AI traders alone make a thin, jumpy book.