(In)Stability Properties of
Limit Order Dynamics
Collaborators:
Eyal Even-Dar (Final Israel)
Sham Kakade (Penn/Wharton Stats)
Yishay Mansour (Tel Aviv University)
[EC 2006]
“Backtesting” of Trading Strategies
• Common microstructure backtesting process:
– assume access to historical limit order data
– reconstruct complete order books at each point in time
– insert hypothetical limit orders into the stream
– simulate forward the execution of the hypothetical orders
• Faithfully simulates the mechanical aspects of market impact
• What about the reactive or “psychological” aspects?
• Formalize as a question about dynamical stability:
– Make various assumptions about how future orders do or do not react to the past
– Can tiny perturbations of the limit order sequence cause dramatic future change?
– Butterfly Effects and Chaos
Two Models of Market Impact
• Both models deal with arbitrary, fixed sequences
• Absolute model:
– market given by a sequence of “absolute” limit order prices (one share each)
– e.g. M = (p_1,buy),(p_2,buy),(p_3,sell),…
– order books constructed from sequence M
– “mechanical” impact only
– motivation:
• traders with “inherent” valuations
• traders with slow time scales, long investment horizons, poor microstructure access
• Relative model:
– market given by a sequence of limit order prices relative to current bid & ask
– e.g. M’ = (d_1,buy),(d_2,buy),(d_3,sell),…
– construct order books & actual prices in concert with each other
• e.g. limit price p_2 = current bid + d_2; limit price p_3 = current ask + d_3; etc.
– crude form of “psychological” or “reactive” impact
– motivation:
• traders “looking for a bargain”; trading off time for price
• “penny-jumping”, optimized execution
• high-frequency traders with low latency and full microstructure access
• How do these models differ?
Stability
• Consider sequences in the two models:
– absolute: M = (p_1,type_1),(p_2,type_2),…
– relative: M’ = (d_1,type_1),(d_2,type_2),…
• Now consider a small, arbitrary modification to each
– e.g. deleting or adding a single order
• (p_i,type_i) from M, (d_i,type_i) from M’
• think of this as “our” action
• How much can such a change alter basic properties of the sequence?
– stability = small change not amplified with time
– instability = small change greatly amplified
• Absolute model: Every “reasonable” property stable!
– volume executed, VWAP, closing price,…
– note: must still be careful; some bounds depend on spread of M
– generalizes to larger modifications, other types
• Relative model: Most properties highly unstable!
– can find sequences (with bounded spread) such that single deletion causes
arbitrarily large changes in volume executed, VWAP, closing price,…
Absolute Model Stability
• <B,S> = original buy and sell books (at some point in simulation)
• <B’,S’> = modified buy and sell books (at the same point)
• Introduce “meta-states” with small “edit distance” between simulations
• E.g. meta-state where B = B’ and S U {s’} = S’ U {s} for some s != s’
• Main technical lemma establishes:
Some Sobering Philosophy
• The “usual” backtesting concern:
– Past strategy performance may not be indicative of hypothetical future performance
• changes in underlying market conditions
• overfitting the historical data
• An even worse concern:
– Past strategy performance may not be indicative of hypothetical past performance!
• well beyond measurable trading costs, mechanical market impact, etc.
• Standard backtesting methodologies implicitly assume an absolute model
– May be fine on longer timescales, but potentially dangerous at microstructure level
– Alternatives: only use actual past trades or live trading
Simulations
0.0025% 0.7%
A Mixture Model
fraction α of absolute traders, 1-α of relative traders, single order deletion
