Rate Cloud

OTC derivative data is available but….

• 100+ fields per trade
  • Different interpretation of field based on asset type
    • Interest rate and credit derivatives
    • Equity derivatives
    • Commodity derivatives
  • Mix of categorical factors
    • Compounding frequency
    • Reset index (LIBOR/Prime/etc)
    • Existence of collateral agreement and terms
    • Traded on/off exchange
    • Cleared/uncleared
• Several factors have non-linear impact

Target: Find a way to quantify what is going on with the market

Define a Standard Instrument and invert its value from recent prices

• Analayze the different terms to identify most common combinations
• Select relevant factors using business knowledge
• Pick model
  • Fit
  • Check errors
  • If not acceptable errors, modify and fit again changing # of points and trees

Model Choices

• Standard regression would be very hard to use due to non-linearity
• Logit - fits non-linear but so many combinations
• Tree method - can handle non-linearity and time dimension

Chosen approach: tree based Random Forest

Used the “caret” package in R

• Keeps track of train/test/validate data subsets automatically
• Has method parRF, which is a fast, parallel thread CPU version of randomForst
• Interfaced with Shiny
• Public data version running on my server at http://ruxton.webhop.net:3839/fkotey/sdr

Code:

irFitTimeGLM <- train(PRICE_NOTATION ~ .,
                    data = irSort[trainSlices,],
                    method = "parRF",allowParallel = TRUE)
pred <- predict(irFitTimeGLM,irSort[testSlices,])
true <- irSort$PRICE_NOTATION[testSlices]
meanErr[k]=mean(pred-true)
meanPcnt[k]=mean(meanErr)/mean(true)
sdErr[k]=sd(pred-true)