Skip to main content
Catalog
M014
Markets

Algorithmic Trading Herding

HIGH(85%)
·
February 2026
·
4 sources
M014Markets
85% confidence

What people believe

Algorithmic trading provides market efficiency and better price discovery.

What actually happens
ImprovedAverage bid-ask spread
+500%Flash crash frequency
-90%Liquidity during stress events
+40%Intraday volatility
4 sources · 3 falsifiability criteria
Context

Algorithmic trading now accounts for 60-70% of US equity volume. The promise was efficiency — algorithms process information faster, reduce spreads, and provide liquidity. And in normal conditions, they do. But most trading algorithms are trained on similar data, use similar signals, and optimize for similar objectives. When market conditions shift, they react simultaneously in the same direction. This creates herding behavior at machine speed. Flash crashes, liquidity vacuums, and cascading sell-offs happen in milliseconds, faster than any human circuit breaker can respond. The 2010 Flash Crash, the 2015 ETF dislocation, and the 2020 COVID crash all showed algorithmic herding amplifying volatility rather than dampening it. The market is more efficient on average but more fragile at the extremes.

Hypothesis

What people believe

Algorithmic trading provides market efficiency and better price discovery.

Actual Chain
Similar algorithms create correlated behavior(60-70% of volume moves in sync during stress)
Momentum signals trigger simultaneous selling
Mean-reversion algorithms withdraw liquidity at worst moments
Crowded trades unwind simultaneously
Flash crashes occur at machine speed(Market drops 5-10% in minutes, recovers in hours)
Stop-loss cascades amplify initial moves
Retail investors hit by price dislocations
Circuit breakers too slow for algorithmic speed
Liquidity becomes illusory(Deep order books vanish during stress)
Market makers pull quotes simultaneously
Bid-ask spreads explode during volatility
Impact
MetricBeforeAfterDelta
Average bid-ask spreadPre-algo baseline-50% (normal conditions)Improved
Flash crash frequencyRareMultiple per year+500%
Liquidity during stress eventsReduced but presentVanishes completely-90%
Intraday volatilityBaseline+40% during stress+40%
Navigation

Don't If

  • Your trading algorithm uses the same signals and data as the majority of market participants
  • Your strategy assumes liquidity will be available during market stress

If You Must

  • 1.Build algorithms that reduce exposure during crowded-trade conditions
  • 2.Implement kill switches that activate before exchange circuit breakers
  • 3.Stress test against flash crash scenarios, not just normal volatility
  • 4.Diversify signal sources to avoid herding with similar algorithms

Alternatives

  • Slower-frequency strategiesTrade on fundamentals over days/weeks, not milliseconds
  • Anti-herding algorithmsStrategies that profit from algorithmic crowding and mean reversion
  • Human-in-the-loop tradingAlgorithmic execution with human oversight for risk management
Falsifiability

This analysis is wrong if:

  • Algorithmic trading reduces market volatility during stress events compared to human-dominated markets
  • Flash crash frequency decreases as algorithmic trading share increases
  • Algorithmic market makers maintain liquidity provision during extreme volatility events
Sources
  1. 1.
    SEC/CFTC Report on the 2010 Flash Crash

    Official investigation documenting how algorithmic trading amplified the May 6, 2010 crash

  2. 2.
    Bank for International Settlements: Algorithmic Trading and Market Quality

    Research showing algorithms improve liquidity in normal conditions but amplify stress

  3. 3.
    Journal of Finance: Algorithmic Trading and the Market for Liquidity

    Academic evidence that algorithmic market makers withdraw during volatility

  4. 4.
    Bloomberg: The Rise of Machine Trading

    Analysis of algorithmic trading's growing share of market volume and systemic risk implications

Related

This is a mirror — it shows what's already true.

Want to surface the hidden consequences of your market assumptions?

Try Lagbase