AI Traders & News¶

Objective¶

Run a swarm of AI traders over a market whose prices have a hidden true value behind them, then move that market with news: trade an earnings surprise, decide what to do with a rumour, watch a retraction unwind, and see how a breakthrough and a factory fire move two rival chipmakers.

Background reading

Prerequisites¶

Chapters 01–13 completed.
The s150-nominal example deployed (Exercise 1). It registers the AI traders AI001–AI020, the student desks TRADER01–TRADER10, the instructor OPS01 (ADMIN) and the market maker MM01, and it ships the market model's market_sim.yaml and the swarm's swarm.yaml.

Background¶

Three processes make the AI market:

pm-market-sim — the market model. Every symbol gets a true value that moves like a company's value does: with the market, with its sector, on its own, and now and then in a jump. Students never see it. It also publishes news: headlines that move the values.
pm-ai-swarm — runs many AI traders, each logged on through pm-alf-gwy as its own participant (AI001, AI002, …) exactly like a student desk. pm-ai-trader runs a single one.
pm-news — the instructor's news desk: publish a headline now, or play a timed scenario.

Each AI trader is a preset: what it trades on (its strategy), how it places orders, how fast, and its risk limits. pm-ai-trader --list-presets prints them all:

Preset	Trades on	Character
`noise-retail`	nothing in particular	Small random orders resting near the touch
`patient-liquidity`	nothing in particular	Small orders waiting deep in the book, giving it depth
`market-taker`, `iceberg-seller`	nothing in particular	Takes liquidity with MARKET orders; hides size behind icebergs
`scalper`, `trend-follower`, `block-taker`	momentum	Buys what has been rising, with IOC, marketable or fill-or-kill orders and protective stops
`contrarian`, `auction-player`	reversion	Fades moves; trades the opening and closing auctions
`value-investor`, `institutional`	the true value	Buys below its own (imperfect) estimate of the value, sells above it
`news-trader`	headlines	Trades a headline within seconds, hard, in its direction

The value and news traders are why prices end up near the values and react to news; the noise and momentum traders are why they never sit exactly on them. A headline has a public sentiment (−1 bad to 1 good) and a secret impact on the value. A rumour carries a credibility and moves no value until it is confirmed; a retracted rumour never moves anything.

Exercise 1: Deploy an Example with AI Traders¶

pm-config-deploy --example s150-nominal

This replaces the configuration you have built since chapter 01. From here on the participants are the ones the example defines — TRADER01–TRADER10, the instructor OPS01 (ADMIN), MM01, and AI001–AI020 — so GW_ADMIN, TRADER03, MM_AAPL_01 and the other gateways of your own configuration no longer exist. Stop the consoles and bots that use them. To return to your own configuration later, deploy engine_config.yaml again.

The scenario and swarm files this chapter uses ship with EduMatcher, but not in your current directory: on a container or pipx install they live inside the installed package, and only in a repository checkout are they under docs/examples. Ask Python where they are, once per terminal, and use $EX in the commands below:

EX=$(python -c "from edumatcher.config_deploy import examples_root; print(examples_root().parent)")
ls "$EX/news"          # the four scenarios: chip-race, earnings-surprise, rumour-confirmed, rumour-retracted

(In the container shell python is the right interpreter; in a Poetry checkout use poetry run python; with pipx use ~/.local/pipx/venvs/edumatcher/bin/python.)

Restart pm-engine (and reopen the market in a moment), then start the gateway the AI traders log on through, the market model, and a market maker for all 150 symbols that quotes around the model's true values, each in its own terminal:

pm-alf-gwy -v
pm-market-sim -v
pm-mm-bot --gateway-id MM01 --all-symbols --strategy passive --gap 0.20 --qty 200 --anchor-sim 0.3

Without --anchor-sim the market maker keeps refilling its quotes around its own mid, and no amount of informed trading moves the price.

Open the market from the operator console:

pm-admin --id OPS01
[OPS01|ADMIN]> SESSION|STATE=PRE_OPEN
[OPS01|ADMIN]> SESSION|STATE=OPENING_AUCTION
[OPS01|ADMIN]> SESSION|STATE=CONTINUOUS

pm-market-sim starts moving the values when continuous trading starts. Check it:

pm-market-sim --status --id OPS01

Checkpoint: pm-market-sim: RUNNING … symbols=150.

Exercise 2: One AI Trader¶

pm-ai-trader --id AI001 --preset noise-retail --symbols AAPL -v

The trader logs on, reads the AAPL book and starts resting small orders near the touch. Watch them arrive with BOOK|SYM=AAPL in the operator console. Stop it with Ctrl+C: its session closes and the participant's disconnect_behaviour (CANCEL_ALL) removes its orders.

Checkpoint: AI001's orders appear in the AAPL book and disappear when you stop it.

Exercise 3: The Swarm¶

pm-ai-swarm --swarm "$EX/ref_data/s150-nominal-setup/swarm.yaml" -v

swarm.yaml gives the twenty traders AI001–AI020 a mix of every preset (per_agent: 150 — every agent trades every symbol, not a slice of them), and sets a budget of about 40 order actions a second for the whole swarm. Once a minute the swarm logs a status line:

[w1-AI001-AI020] t=120s session=CONTINUOUS actions=4801 (40.0/s) live_orders=164 fills=1290 rejects=0

Checkpoint: a status line every minute; books changing on symbols nobody in the room is trading.

Exercise 4: Price and Value¶

The instructor can see what the students cannot:

pm-market-sim --status --id OPS01 --top 5

symbol        value        mid  mid/value
TSLA         212.40     219.95     +3.55%
...

The value traders lean against these gaps; the noise and momentum traders keep opening new ones. Run the command again a few minutes later: the symbols at the top change, and large gaps rarely last.

Checkpoint: mid/value gaps of a few percent at most, changing from one look to the next.

Exercise 5: An Earnings Surprise¶

Every student opens a news display: the News screen of the trading GUI, the News view of TapeDeck, pm-ticker, or simply

pm-news list

Then the instructor plays the first scenario:

pm-news --id OPS01 play "$EX/news/earnings-surprise.yaml"

Sixty seconds into the play (it counts from its start when the market is already in continuous trading) AAPL reports far above expectations. The news is confirmed from the start: AAPL's value jumps about 8% at once. The news traders hit the book within seconds; the value investors follow.

Students try to buy AAPL as soon as they see the headline:

[TRADER01]> NEW|SYM=AAPL|SIDE=BUY|TYPE=MARKET|QTY=100

Compare your fill with the AAPL price a minute later.

Checkpoint: AAPL trades several percent higher within a minute of the headline.

Exercise 6: A Rumour That Turns Out True¶

First stop the market maker (Ctrl+C on pm-mm-bot; its quotes are withdrawn). It quotes around the true values, and a rumour changes no value: it would sell to every rumour buyer at the old price, and the rumour would leave no trace. Without it, prices are what the traders believe — the AI traders' resting orders keep the books two-sided.

pm-news --id OPS01 play "$EX/news/rumour-confirmed.yaml"

At 60 seconds a rumour says JPM is a takeover target, credibility 50%. Nothing has happened to JPM's value — any move now is traders betting on the rumour. Discuss before the next step: buy now, or wait? At 240 seconds the bid is confirmed and the value jumps about 12%.

Checkpoint: JPM moves a little on the rumour and much more on the confirmation.

Exercise 7: A Rumour That Is Retracted¶

pm-news --id OPS01 play "$EX/news/rumour-retracted.yaml"

At 60 seconds a rumour says energy producers face a new windfall tax: a sector headline, so every ENERGY symbol (XOM, CVX, …) is concerned. The values do not move; prices dip only because traders believe it. At 240 seconds it is retracted — in the GUIs the rumour is now struck through — and prices drift back to where the values are.

Checkpoint: ENERGY symbols dip after the rumour and are back near their earlier prices a few minutes after the retraction.

Exercise 8: A Breakthrough and a Fire¶

Not all news is earnings. Events (a fire at a factory, a research breakthrough, a big new contract, a product recall) go one way only, and they are big. Their kind says what happened: OPERATIONS, RESEARCH, CONTRACT, ESG or PRODUCT. Keep the market maker stopped, and have students watch NVDA and AMD side by side.

pm-news --id OPS01 play "$EX/news/chip-race.yaml"

At 60 seconds NVDA announces a RESEARCH breakthrough: its next chips are 50% faster. The news is confirmed, and NVDA's value jumps about 10%. It is a symbol headline, so AMD's value does not move at all. Discuss: should AMD's price fall because its rival got ahead?

At 240 seconds a rumour says AMD's main chip fab is on fire, credibility 70%. Discuss before the next step: sell AMD now, or wait for confirmation? A credibility of 0.7 is high, and a fire never turns out to be good news. At 360 seconds the fire is confirmed and AMD's value drops about 7%.

Checkpoint: NVDA jumps on the first headline while AMD barely moves; AMD dips on the rumour and falls further on the confirmation.

Restart the market maker afterwards (the command from Exercise 1).

Exercise 9: Stop the Swarm¶

Ctrl+C on pm-ai-swarm stops every worker; each closes its traders' sessions, which cancels their orders. Stop pm-market-sim with Ctrl+C too: it saves its values and continues from them next time.

Checkpoint: no AI0xx orders left in the books; the market maker's quotes remain.

Classroom Tips¶

Your own headlines need no scenario: pm-news --id OPS01 inject --symbol TSLA --kind PRODUCT --sentiment -0.6 --impact -0.05. A rumour you inject (--rumour --credibility 0.4) waits for your pm-news --id OPS01 confirm N7 or retract N7 — the id is printed when you inject it.
pm-market-sim also generates news on its own, a handful a day. Set the rates in market_sim.yaml to 0 for a class where only your headlines should appear.
Same swarm.yaml and market_sim.yaml seeds, same traders and same values: change seed in either file for a different market.
A larger --anchor-sim (up to 1) makes the market maker follow the true value faster; then the AI traders, and the students, have less of a gap to trade.
Running the AI-Trader Swarm covers sizing, logs and daily summaries for longer runs.

Reflection¶

Why does a confirmed headline move the price further than a rumour with the same sentiment? Who in the swarm was on the other side of the news traders' orders right after the earnings headline? The market maker quotes around the true value: what would have happened to the price, and to the students' chances of buying AAPL cheaply, if it had not been anchored? An earnings headline can go either way, but a fire can only be bad: how did that change what you did on the AMD rumour?