Information Cascades and Herd Behavior

Prices sometimes move not because thousands independently updated beliefs, but because traders assume the previous trader knew something. That dynamic—information cascades—turns prediction markets into social proof machines, for better and worse.

This chapter connects strategic play, manipulation costs, arbitrage, and news-driven jumps elsewhere in this module.

Private signals vs public actions

Each trader may receive:

• Private signal — own model, insider whisper (legal or not), local gossip
• Public history — prior trades, price path, volume spikes

In a cascade, public history swamps private signal. Trader i ignores their own 48% model because trader i-1 bought aggressively—inferring i-1 was informed.

Formal cascade intuition (binary)

Simplified Anderson–Burch sequence:

1. Traders act in order.
2. Each sees predecessors’ actions, not their private signals.
3. If enough early actors bought YES, later actors buy YES even when their signal said NO.

Outcome: unanimous buying with weak aggregate information if early movers were noise.

When cascades are rational

Early traders might be:

• Informed — leak, poll aggregator, court filing
• Noise — meme, mistaken click, manipulation

Bayesian updating on action without knowing type creates herding equilibrium.

If you cannot distinguish informed from noise, following the herd minimizes regret—game theory “avoid being lone idiot.”

Prediction market amplifiers

Crypto venues with wallet labels intensify copy trading: “whale 0x… bought $50k YES.”

Herd vs wisdom of crowds

As in the chapter on when crowds beat experts, crowds work with diverse independent signals.

Cascades destroy independence—everyone conditions on same price chart.

Test: Did price move with volume from diverse timestamps or one burst?

Bubble dynamics within [0,1]

YES at 15% → 35% → 55% in an hour without news is often cascade + manipulation budget, not new fundamentals.

Fade trade game:

• Faders bet reversion when no arb to other venues
• Momentum traders ride cascade
• Equilibrium uncertain until news or arb

Stopping cascades

Arbitrage — if related contract still 40%, arbs sell expensive YES.

Market makers — widen spread, refuse to lift.

Platform design — delay displaying % until $X volume; show confidence interval from depth.

Informed late movers — large counter-trade when cascade overshoots model.

Informational vs payoff cascades

Informational — believe others know more.
Payoff — career risk of being wrong alone (analyst herd on TV).

Prediction markets blend both: money plus Twitter visibility.

Case pattern: debate night

T+2 min — Candidate A zinger clip viral.
T+5 min — Retail buys A YES on three apps.
T+8 min — Price 52% → 61% on thinnest venue.
T+15 min — Pollsters say clip within margin; Kalshi still 54%.
T+40 min — Arb compresses thin venue to 55%.

Cascade overshoot on thin leg; thick leg moderates “truth.”

1. Pre-commit model — write probability before tape.
2. Ignore % for X minutes after vertical move.
3. Check composite across venues.
4. Ask for news object — tweet ≠ filing.
5. Size down when move / volume ratio extreme.

Building narratives without trading

Journalists citing single venue after spike amplify cascade—ethical reporting uses liquidity-weighted composite and mentions thin market caveat.

Connection to manipulation

Manipulators seed early buyers to trigger cascades, lowering cost if others finish the pump. Defense combines arbitrage and transparency.

• Cascades trade on actions, not only facts.
• They break independence required for crowd wisdom.
• Thin markets + social proof = overshoot risk.
• Arb, delays, and composite pricing are antidotes.

What comes next

Next: How news enters the game and triggers price jumps.