TODO

This document contains some notes about my TODO list of possible advanced capabilities to be added to PortfolioAnalytics. It is neither a promise nor a project plan, but rather some ideas of features and direction that I think would be additive to the tool set and the field of portfolio construction decision making.

asset selection cross validation

• what impact does a different asset selection have on the optimization?
• allow user to do sequential optimizations leaving n assets out of the optimization
• sample() should be able to construct the different portfolios
• construct a comparison report on top neighbor solutions

OOS degradation

• in backtesting, you want to measure how the OOS strategy performed vs. the expectation
• build similar tools for portfolio construction
• evaluate rebalanced portfolios based on how they perform out of sample, and how the optimizer would choose differently on the next rebalancing period
• how do we make portfolio selection more robust out of sample? (see neighbors discussion below)

tools to aid selection among neighbors in multi-objective optimization

• the top n portfolios in multi-objective optimizations are balancing several competing objectives
• multiple portfolios will show different trade offs
• user preferences may be weakly dominant, and so the “best” portfolio based on equal or slightly weighted objective may not really be the best available solution, especially out of sample.

integrated storage of trials

• storing trials is important to evaluating multiple testing bias
• also important for reproducibility if results used in reporting
• allow for a trials/ directory in a project, and experiment notes about the trial
  • typically reason and short_reason are what I use
  • store assets, returns, portfolio object, optimization output
  • construct a meaningful filename, e.g. paste0('trials/results_trial_',format(Sys.time(),'%F_%R'),'.',reason_short,'.rda')
  • I have historically stored multiple variables, should this store an environment instead so you can examine multiple trials?

integrate mcsim and other confidence interval methods

• Monte Carlo analysis and influence function analysis can answer question about confidence intervals for returns and risk metrics on a generated portfolio
• returns of portfolio could be processed by mcsim() or similar

multiple solvers in portfolio sets

• store preferred solver in portfolio spec
• benchmark portfolios like min variance or min ES or Markowitz portfolio could be solved quickly
• target portfolios with more complex objectives could use DEoptim, random, or mco

hyper parameter tuning

• choice of solver parameters can vastly change speed and quality of solutions for global solvers
• build tools for tuning hyper parameters of each solver
• build tools for comparing different solvers
• include convergence charts for different solvers and hyper parameters

portfolios of unequal length

• assets start and stop being available to trade all the time
• add support for either assets with longer /shorter histories
• and support for assets which close
• and support for assets that cease existence
• shrink the available universe and maintain things like full investment constraint
• support edge case of n=1

FMMC integration