Uncovering Hidden Structure in Bond Futures Trading
Fei Chen, Stephen Figlewski, Jeffrey Heisler, Andreas S. WeigendInformation Systems Department, Stern School of Business, NYU
Outline
- Problems
- Data
- Feature Vector
- Trade-specific variables
- Market variables
- Method
- Idea
- Characterize by prototypes
- Model
- Optimization
- Expectation-Maximization (EM) Algorithm
- Results
- Static
- Distribution of accounts across clusters
- Dynamics
- Transition of accounts between clusters
- Return and Risk
- Future Directions
Problems
- Behavioral Finance in Bond Futures Market
- What are the underlying trading styles?
- Do we find behavior consistent with technical trading?
- How consistently do speculators adhere to a trading style?
- Can we predict transitions from one style to another?
- What are profit and risk characteristics for these styles?
Data
- Treasury Bond futures Market
- Large trading activity
- “Closed universe”
- Commodity Futures Trading Commissions (CFTC)
- September 1992 contract: 3 million transactions
- Each transaction consists of
- Contract
- Date
- Time
- Price
- Buy/Sell
- Volume
- Customer Type Indicator (CTI)
- Account identification
Example
9209 1992 06 02 13 45 99060 S 4 3 0000011800 596*
9209 1992 06 02 11 32 99060 B 1 4 1007121118 766*
9209 1992 06 02 12 37 99060 B 10 4 1677*UK501 235*
9209 1992 06 02 12 37 99060 S 10 3 0000020251 333*
9209 1992 06 02 7 21 99060 B 3 4 0383*50005 110*
9209 1992 06 02 7 21 99060 S 3 4 0385*56000 110*
9209 1992 06 02 7 22 99060 B 2 4 0382*50005 110*
9209 1992 06 02 7 22 99060 S 2 3 0000089907 663*
9209 1992 06 02 12 37 99060 S 2 4 12502+++++ 056*
9209 1992 06 02 12 37 99060 S 2 4 12502+++++ 056*
9209 1992 06 02 12 37 99060 B 1 4 065860295+ 616*
9209 1992 06 02 12 37 99060 S 1 4 0114043661 476*
9209 1992 06 02 12 37 99060 B 1 4 065860295+ 616*
Trade
Constructed from individual records
Features (1): Trade-specific
Features (2): Market-Specific
- Relative price
- At opening
- At time of first reversal
Outline
- Problem
- Data
- Method
- Clustering
- Results
- Future Directions
Outline
- Problem
- Data
- Method
- Clustering
- Results
- Future Directions
Outline
- Problem
- Data
- Method
- Clustering
- Results
- Future Directions
Outline
- Problem
- Data
- Method
- Clustering
- Results
- Future Directions
To estimate the parameters, EM
Clustering
- Data Generation Model
- Mixture models
- Data generating process
- Data Recognition Model
- Number of clusters known
- Gaussian noise
- Structure of covariance matrix for each cluster
Results
- Conditional Distributions
- Trade-specific variables
- Trade length
- Maximum position size
- Market-specific variable
- Relative price at opening, and at first reversal
- First look at static information
-
7 Clusters
1 2 3 4 5 6 7
size 2465 1969 1592 1265 980 599 401
PPT Slide
PPT Slide
PPT Slide
97% Short 88% Long 71% Long
Dynamics
- Questions: What is the probability that an account moves from cluster j to cluster k in the next trade?
From ?
To 1 2 3 4 5 6 7
7 22 51 50 66 34 21 124
6 85 130 80 116 60 40 0
5 148 160 191 128 161 70 0
4 112 190 173 286 142 68 1
3 241 296 370 211 185 127 1
2 449 534 272 191 166 115 4
1 1257 438 248 105 125 96 4
PPT Slide
Profit
-
-
-
N mean error mn/er
7 401 6330 5930 1.06
6 599 160 470 0.34
5 980 550 370 1.50
4 1265 -9620 6990 1.37
3 1592 390 180 2.17
2 1969 -400 80 5.37
1 2465 6 40 0.17
Iterative nature of modeling
- Revisit assumptions
- Number of clusters
- Covariance structure
- Sanity checks
- Different algorithms
- Include additional randomized variables as inputs
- Entropy
Iterative nature of modeling
- Revisit assumptions
- Number of clusters
- Covariance structure
Accounts
- Question: Take the unconditional probabilities of the clusters, how non-random are the accounts distributed
Conclusions
- Unsupervised learning
- Make assumptions
- trade style assumptions
- statistical assumptions
- Reveal structures in the data
- Day trades vs longer trades
- L/S and relative price
- Profit
- Persistent in one trading style
- Validate findings
- Different assumptions
- Different algorithms
- Randomization
- Entropy
Next step
- More data
- More variables
- Technical indicator
- Information related to the evolution path of trades
- Compare CTI types
Faces
log.exposure area of face
rel.open.mov shape of face
log.chro.length length of nose
log.max.cumsum location of mouth
log.ent.volat.30 location of eye
ent.price.30 shape of eye
ent.till location of pupil
ent.vlm.3000 angle of eyebrow
% at night width of eyebrow
Further information
Fei Chen, Andreas S. Weigend
Information Systems Department
Stern School of Business / NYU
44 West Fourth Street, MEC 9-74
New York, New York 10012
Phone: +1 (212) 998-0803 (Chen)
Phone: +1 (212) 998-0803 (Weigend)
Fax: +1 (212) 995-4228
fchen@stern.nyu.edu
aweigend@stern.nyu.edu
www.stern.nyu.edu/~aweigend
Stephen Figlewski
Salomon Center
New York University
44 West Fourth Street, MEC 9-64
New York, New York 10012
Phone: +1 (212) 998-0712
Fax: +1 (212) 995-4220
sfiglews@stern.nyu.edu