News and Rumours

Learning objectives

After reading this chapter you will be able to:

• Read a headline the way the simulation means it: scope, kind, sentiment and the secret impact behind it
• Follow a rumour from its first headline to its confirmation or retraction
• Predict how the AI traders and the market maker react to news
• Turn generated news up, down or off, and publish news of your own in front of a class

Prerequisites: The True Value.

Why news

In the simulation, as on a real exchange, the true value of a share does two things: it drifts a little all the time, and now and then it jumps because something happened. The jumps are what news is for. A headline tells the market that the value has moved — and roughly which way — before the price has caught up. Who reads it first, who believes it and who ignores it is what turns a headline into a price move you can watch.

Anatomy of a headline

Field	Public?	Meaning
Scope and targets	yes	One symbol, one sector, or the whole market
Kind	yes	EARNINGS, GUIDANCE, MNA, REGULATORY, PRODUCT, LEGAL, MANAGEMENT, MACRO, OPERATIONS, RESEARCH, CONTRACT or ESG
Headline	yes	The text, e.g. "NVDA raises its outlook"
Sentiment	yes	How good it sounds, from −1 (very bad) to +1 (very good)
Status	yes	CONFIRMED (it happened), RUMOUR, then CONFIRMED or RETRACTED
Credibility	yes, rumours only	How believable the rumour looks, 0 to 1
Impact	no	The log move it causes in the true value: 0.05 is about +5.1 %, −0.10 about −9.5 %

Students see everything but the impact: in terminal-gui's news view, the API (GET /api/v1/news) and pm-ticker. The impact is the market model's secret, exactly like the true value.

How far a headline moves values. A symbol headline moves that symbol by its impact. A sector headline moves each symbol of the sector by beta_sector × impact, a market headline every symbol by beta_market × impact — so with the s300-sim betas a sector headline moves its symbols by about 35–60 % of the impact, and a market headline every symbol by 35–60 % of it.

Generated news

Left alone, pm-market-sim writes news itself, at random moments during continuous trading. The s300-sim example keeps the defaults:

news:
  rate_symbol: 6.0         # symbol headlines per trading day, whole exchange
  rate_sector: 0.5         # sector headlines per trading day
  rate_market: 0.3         # market headlines per trading day
  rumour_share: 0.3        # share that starts as a rumour
  rumour_confirm: 0.6      # chance a rumour is confirmed rather than retracted
  rumour_delay_min: 0.02   # trading days before a rumour resolves (uniform)
  rumour_delay_max: 0.15

What that means on a 300-symbol exchange:

	On average
Symbol headlines	6 a day — one every hour or so; any one share gets news about once every 50 trading days
Sector headlines	one every two trading days
Market headlines	one every three or four trading days
Rumours	about 2 a day; they resolve 8 to 60 minutes of continuous trading later, 3 in 5 confirmed

A "trading day" here is the 6.5 hours of continuous trading; the rates are per trading day whatever speed the calendar runs at (see Accelerated Time).

What kind of news, how big

Each headline is drawn from a catalogue of stories, and each story has its own size: the impact is normally distributed with the story's mean and spread. There are two sorts.

Everyday news can go either way. Its headline follows the sign of the impact: "NVDA beats earnings expectations" or "NVDA misses earnings estimates".

Kind	Scopes	Share of symbol news	Impact mean	Impact spread (σ)	Typical move
EARNINGS	symbol	24 %	0	5 %	±5 %
GUIDANCE	symbol	12 %	0	3 %	±3 %
PRODUCT	symbol	12 %	+0.5 %	2 %	±2 %
REGULATORY	symbol, sector	6 %	−1 %	4 %	±4 %, a little more often bad
LEGAL	symbol	6 %	−1 %	3 %	±3 %, a little more often bad
MANAGEMENT	symbol	6 %	0	2 %	±2 %
MNA	symbol	3 %	+6 %	6 %	a takeover bid: mostly up, sometimes a lot
MACRO	sector, market	—	0	1.5 %	±1.5 % before the betas

Events go one way only, and they are big: a fire is always bad news, a breakthrough always good. Together they are about a third of symbol news, so with the default rates a class sees two or three a day, which is far more often than in real life, so that students get to watch what such a shock does to a price.

Kind	Examples	Typical move
OPERATIONS (15 %)	"Fire ravages XOM's main refinery", flooding, a cyber-attack, a strike, a parts shortage; good: a site reopening early, a new oil find	−2.5 % to −5 %; +2 % to +5 %
ESG (7 %)	"Environmental protests delay TSLA's new battery plant by 12 months", a fine for breaking environmental rules, an oil spill, a wildfire blamed on a utility's lines; good: a permit won after a long battle	−2 % to −8 %; +2.5 %
CONTRACT (5 %)	"CAT signs a $5 bn long-term deal with a major new customer", losing the largest customer, a defence contract	+4.5 % to +5 %; −5 %
RESEARCH (4 %)	"NVDA breakthrough lifts next-generation chip performance by 35 %", a late-stage drug trial that succeeds or fails, a flagship R&D project written off, a chip that slips a year	+5 % to +8 %; −5 % to −10 %
PRODUCT (2 % of the 14 %)	"GM recalls 300,000 vehicles over a brake defect", a product recalled over a safety defect, deliveries halted	−6 %

Some events only happen in their own industry: chip stories to chipmakers, trials to drug makers, recalls of vehicles to carmakers, groundings to airlines, defence contracts to aerospace. Other events need a fitting sector: a bank has nothing to recall, and a utility has no largest customer. A number in the headline grows with the impact: a chip that is 40 % faster moves the value more than one that is 20 % faster.

Sector headlines are REGULATORY (42 %, including export curbs on chips and drug-pricing reform), MACRO (40 %) or OPERATIONS (18 %: a supply-chain disruption, a port strike, a hurricane that shuts refineries). Market headlines are always MACRO.

Sentiment: a hint, not the answer

The sentiment students see is the impact made public, but blurred: for everyday news roughly impact ÷ (2 × the story's spread), plus a little noise (σ 0.15), clipped to ±1. So an earnings beat of one spread (+5 %) reads about +0.5, and a big miss of two spreads (−10 %) about −1. For an event it is impact ÷ (|mean| + spread) plus the same noise, so a typical fire reads about −0.7 and a bad one −1. An event's sentiment always has the right sign. For everyday news the sign is almost always right. For both, the size is only a guide. Two headlines with the same sentiment can move the value by quite different amounts, which is why value traders, who see the value itself, keep correcting what news traders start.

Rumours

stateDiagram-v2
    [*] --> RUMOUR: headline with a credibility\nvalue does not move
    RUMOUR --> CONFIRMED: confirmed\nvalue moves by the impact
    RUMOUR --> RETRACTED: retracted\nvalue never moves
    CONFIRMED --> [*]
    RETRACTED --> [*]

A rumour is a headline that may not be true. It carries a credibility and moves nothing in the model. Later it is confirmed — a second headline, and only then does the value move — or retracted, and the value never moves at all. Prices often do move on a rumour, though: that is traders betting on it, and when a retracted rumour's price move unwinds, someone has lost money.

Credibility is informative but not reliable. A generated rumour that will be confirmed gets a credibility between 0.4 and 0.9; one that will be retracted between 0.1 and 0.6. A credibility of 0.8 is almost surely true, one of 0.2 almost surely false, and anything in 0.4–0.6 is a coin toss.

Who reacts, and how

Who	Reacts to	How
news-trader (5 agents)	every headline touching its symbols	Hears it after a lag of its own (median 5 s), then trades in the headline's direction while its conviction — |sentiment|, times credibility for a rumour, halving every 2 minutes — stays above 0.15; pays up to 0.08 × conviction beyond the price it first saw
value-investor, institutional (30 agents)	the true value, which news moves	Buy or sell toward the new value as soon as it moves. On a rumour, which does not move the value, they shift their own view by 0.05 × sentiment × credibility
MM01, the market maker	the true value	Moves its own price 30 % of the way toward the value every second and re-quotes once that has moved 15 ticks — so a confirmed headline moves its quotes within seconds, and a rumour does not move them at all
Everyone else (65 agents)	prices only	Trend followers and scalpers chase the move once it shows; contrarians fade it

Put together, a confirmed headline is priced in fast: in a test with an anchored market maker and the same kinds of trader on 150 symbols, a +10 % earnings surprise moved the price +6.5 % within 5 seconds and +10 % within 15 seconds. Rumours leave much less trace than you might expect. The market maker sells to every rumour buyer at its value-anchored price, so in the same test a credible good rumour (sentiment 0.7, credibility 0.5) lifted the price under 3 %, and a sector rumour moved its sector by less than 1 %. To show students how a rumour moves a market, run without the market maker — see Tuning the Traders.

Steering the news

Fewer, more, or none

Change the rates in market_sim.yaml (how to deploy your copy: Tuning the Traders):

You want	Set
A quiet market where you alone publish news	every rate_* to 0
A news-heavy market, a headline every few minutes	rate_symbol: 60 (and perhaps rate_sector: 3)
More rumours, harder to read	rumour_share: 0.6, rumour_confirm: 0.5
Rumours that linger for a lesson	rumour_delay_min: 0.1, rumour_delay_max: 0.3
No market-wide shocks	rate_market: 0

Each rate is a Poisson rate: the gaps between headlines are random, so rate_symbol: 60 gives ten an hour on average, sometimes three in a minute, sometimes none for twenty.

Publishing news yourself

pm-news sends a headline to the running market model as an ADMIN participant. In the container:

# From the host
docker exec edumatcher-sim pm-news --id OPS01 inject --symbol NVDA \
    --kind EARNINGS --sentiment 0.8 --impact 0.08 \
    --headline "NVDA smashes estimates on data-centre demand"

# From within the container
pm-news --id OPS01 inject --symbol NVDA \
    --kind EARNINGS --sentiment 0.8 --impact 0.08 \
    --headline "NVDA smashes estimates on data-centre demand"

The value of NVDA moves +8.3 % at once; the price follows within seconds. A rumour you inject waits for you:

# From the host
docker exec edumatcher-sim pm-news --id OPS01 inject --sector ENERGY \
    --kind REGULATORY --sentiment -0.7 --impact -0.06 --rumour --credibility 0.6 N12
docker exec edumatcher-sim pm-news --id OPS01 confirm N12     # or: retract N12

# From within the container
pm-news --id OPS01 inject --sector ENERGY \
    --kind REGULATORY --sentiment -0.7 --impact -0.06 --rumour --credibility 0.6 N12
pm-news --id OPS01 confirm N12     # or: retract N12

inject prints the headline's id (N12 here). A rumour you inject never resolves on its own. Your headlines and generated ones mix freely; set the rates to 0 if yours should be the only ones.

Scenarios

For a lesson, script the news in a file and play it:

# From the host
docker exec edumatcher-sim pm-news --id OPS01 play \
    /opt/edumatcher/.venv/lib/python3.14/site-packages/docs/examples/news/earnings-surprise.yaml

# From within the container
pm-news --id OPS01 play \
    /opt/edumatcher/.venv/lib/python3.14/site-packages/docs/examples/news/earnings-surprise.yaml

Four scenarios ship with EduMatcher — earnings-surprise, rumour-confirmed, rumour-retracted and chip-race (a chip breakthrough, then a rumoured fire at a rival's fab) — and fit s300-sim as they fit the s150-* examples. Their format, and how to write your own, is in The Market Model; the AI Traders & News exercises use them in class.

Watching the news

Where	What
terminal-gui, news view	Every headline as it is published, with status and sentiment
docker exec edumatcher-sim pm-news list	Headlines printed as they arrive, until Ctrl+C
docker exec edumatcher-sim pm-news past --limit 20 --date 2026-10-14	Recent headlines from that session date, newest first
GET /api/v1/news (any API key)	Today's headlines, and the sector of every symbol

Try it

1. Pick a quiet symbol in book-gui and note its price and the value (pm-market-sim --status --id OPS01 lists it if it is far off; the admin API, GET /api/v1/admin/sim with the OPS01 key, gives all of them).
2. Inject a confirmed headline with impact 0.05 and watch the price for a minute. Who traded first? (The tape and the order book show AI ids.)
3. Inject a rumour with sentiment 0.8, credibility 0.9, and no impact worth speaking of (--impact 0.001). Does the price move? Why not much?
4. Retract it. Then try again without the market maker (Tuning the Traders shows how to remove it).

See also

• The Market Model — the reference for pm-news and scenarios
• How the AI Traders Decide — the news and value strategies in full
• AI Traders & News — classroom exercises with news