KUMARAGANESH SUBRAMANIAN

1
MSE 444 Investment PracticeShort and long-term
prediction combination

• KUMARAGANESH SUBRAMANIAN
• XIAOLONG TAN
• PRABAL TIWAREE
• DIMITRIOS TSAMIS
• JUNE 3, 2009

2
Returns Model
3
Using multiple predictors

• Assume that alphas are a linear combinations of
  factors
• Estimate B using pooled panel regression
• Moreover,
• is a positive definitive matrix of
  mean-reversion coefficients

4
Transaction Costs

• Trading shares costs
• Assume that

5
Optimization Problem

• Find the optimal portfolio at each time step by
  solving the following problem
• Use Dynamic Programming!

6
Main result

• Optimal portfolio is linear combination of
  previous position and a moving target portfolio
• where
• and

7
Simplification

• If then

8
Static model

• Solve
• ie fully discount the future
• Solution

9
Experiments

• Use 6 different commodities futures from London
  Metal Exchange
• Evaluate based on gross and net SR and cumulative
  returns
• Compare optimal, static and no TC strategies
• Predictors normalized averages over 5 days, 1
  year and 5 years

10
Cumulative Returns
11
Sharpe Ratios

• Dynamic strategy 0.4707
• Static strategy 0.4618

12
Effect of lambda
13
Rebalancing costs
14
Experiments with shares

• Use predictors provided by EvA
• Short-term stat-arb daily predictors
• Long-term EMN monthly predictors
• interpolate daily values
• There were 1089 securities common across all data

15
Reduce the size of the portfolio!

• Using all the securities produces bad results
• Sis essential to the model, but the quality of
  the estimator deteriorates as the number of
  securities increases
• To evaluate the model try random portfolios and
  observe their performance

16
Using all securities
17
Cumulative Returns with 20 securities
18
Cumulative Returns with 100 securities
19
Cumulative Returns with 500 securities
20
Best portfolio size 19 securities
21
Evaluate based on SR
22
Conclusions

• The strategy works better on commodity data
• The strategy appears to be self-financing
• The strategy does not work well on very large
  portfolios (probably due to parameter estimation
  errors)