SlideShare a Scribd company logo
Experiences with using
        R in credit risk
                   Hong Ooi
Introduction

• A bit about myself:
    • Did an actuarial degree, found it terribly boring and switched to
      statistics after graduating
    • Worked for several years as a general insurance risk modeller
    • Switched to banking for a change of scenery (at the top of the boom)
    • Now with ANZ Bank in Melbourne, Australia

• Contents:
   • LGD haircut modelling
   • Through-the-cycle calibration
   • Stress testing simulation app
   • SAS and R
   • Closing comments




                                     Page 2
Mortgage haircut model

• When a mortgage defaults, the bank can take possession of the property
  and sell it to recoup the loss1
                      p
• We have some idea of the market value of the property
• Actual sale price tends to be lower on average than the market value (the
  haircut)2
• If sale price > exposure at default, we don’t make a loss (excess is passed
       l    i                 t d f lt    d ’t    k    l    (       i       d
  on to customer); otherwise, we make a loss

  Expected loss = P(default) x EAD x P(possess) x exp.shortfall
    p              (       )          (p      )     p




Notes:
1. For ANZ, <10% of defaults actually result in possession
2. Meaning of “haircut” depends on context; very different when talking
   about, say, US mortgages


                                    Page 3
Sale price distribution
                          p


                                                  Valuation




Expected shortfall




                                     Haircut




                                     Exposure
                                     at default




                                $

                            Page 4
Stat modelling

• Modelling part is in finding parameters for the sale price distribution
    • Assumed distributional shape, eg Gaussian
    • Mean haircut relates average sale price to valuation
    • Spread (volatility) of sale price around haircut
• Once model is found, calculating expected shortfall is just (complicated)
  arithmetic




                                     Page 5
Valuation at origination                                       Valuation at kerbside                                            Valuation after possession
                 15




                                                                                15




                                                                                                                                                 15
                 14




                                                                                14




                                                                                                                                                 14
                 13
                  3




                                                                                13
                                                                                 3




                                                                                                                                                 13
                                                                                                                                                  3
log sale price




                                                               log sale price




                                                                                                                                log sale price
                 12




                                                                                12




                                                                                                                                                 12
                 11




                                                                                11




                                                                                                                                                 11
                 10




                                                                                10




                                                                                                                                                 10
                 9




                                                                                9




                                                                                                                                                 9
                      11        12           13      14   15                         10   11      12             13   14   15                         10        11      12           13   14   15

                                     log valuation                                               log valuation                                                       log valuation




                                                                                                Page 6
Volatility

Volatility of haircut (=sale price/valuation) appears to vary systematically:


                Property type                         SD(haircut)*

                A                                         11.6%

                B                                         9.3%
                                                          9 3%

                C                                         31.2%



                 State/territory                      SD(haircut)*
                 1                                        NA

                 2                                       13.3%

                 3                                        7.7%

                 4                                        9.2%

                 5                                       15.6%

                 6                                       18.4%

                 7                                       14.8%

                 * valued after possession




                                             Page 7
Volatility modelling

• Use regression to estimate haircut as a function of loan characteristics
• Use log-linear structure for volatility to ensure +ve variance estimates
      log linear
• Data is more heavy-tailed than Gaussian (even after deleting outliers), so
  extend to use t distribution
   • df acts as a shape parameter, controls how much influence outlying
     observations h
       b      ti    have

• Constant volatility/Gaussian error model is ordinary linear regression
• Varying-volatility models can be fit by generalised least squares, using gls
  Varying volatility
  in the nlme package; simpler and faster to directly maximise the Gaussian
  likelihood with optim/nlminb (latter will reproduce gls fit)
• Directly maximising the likelihood easily extended to t case: just replace all
  *norm functions with *t
    norm                t




                                      Page 8
Normal model residuals




          0.7




                                                                                    6
          0.6




                                                                                    4
          0.5




                                                                Sample quantiles

                                                                                    2
          0.4
Density

          0.3




                                                                                    0
          0.2




                                                                                    -2
          0.1




                                                                                    -4
          0.0




                -4   -2         0            2    4     6                                -3   -2       -1       0        1        2   3

                              Std residual                                                         Theoretical normal quantiles


                                                 t5-model residuals
           .4
          0.




                                                                                    10
          0.3




                                                                Sampl e quantiles

                                                                                    5
Density

          0.2
          0




                                                                                    0
          0.1




                                                                                    -5
          0.0




                -5        0             5          10                                    -6   -4       -2       0        2        4   6

                              Std residual                                                            Theoretical t-quantiles

                                                            Page 9
Example impact

Property type = C, state = 7, valuation $250k, EAD $240k

                                 Gaussian model

         Mean formula       Volatility formula         Expected
                                                      shortfall ($)
         ~1
          1                 ~1
                             1                             7,610
                                                           7 610
         ~1                 ~proptype                      23,686
         ~1                 ~proptype + state              29,931


                                   t5 model

         Mean formula       Volatility formula         Expected
                                                      shortfall ($)
         ~1                 ~1                             4,493
         ~1                 ~proptype                      10,190
         ~1                 ~proptype + state               ,
                                                           5,896


                                    Page 10
Notes on model behaviour

• Why does using a heavy-tailed error distribution reduce the expected
  shortfall?
    • With normal distrib, volatility is overestimated
       → Likelihood of low sale price is also inflated
       • t distrib corrects this
    • Extreme tails of the t less important
       • At lower end, sale price cannot go below 0
       • At higher end, sale price > EAD is gravy
    • This is not a monotonic relationship! At low enough thresholds
                                                             thresholds,
      eventually heavier tail of the t will make itself felt

• In most regression situations, assuming sufficient data, distributional
  assumptions (ie, normality, homoskedasticity) are not so critical: CLT
  comes into play
• Here, they are important: changing the distributional assumptions can
  change expected shortfall by big amounts
      g                           g


                                     Page 11
In SAS

• SAS has PROC MIXED for modelling variances, but only allows one grouping
  variable and assumes a normal distribution
• PROC NLIN does general nonlinear optimisation
• Also possible in PROC IML

• None of these are as flexible or powerful as R
• The R modelling function returns an object, which can be used to generate
  predictions, compute summaries, etc
• SAS 9.2 now has PROC PLM that does something similar, but requires the
  modelling proc to execute a STORE statement first
    • Only a few procs support this currently
    • If you’re fitting a custom model (like this one), you’re out of luck




                                    Page 12
Through-the-cycle calibration

• For capital purposes, we would like an estimate of default probability that
  doesn’t depend on the current state of the economy
              p                                        y
• This is called a through-the-cycle or long-run PD
    • Contrast with a point-in-time or spot PD, which is what most models will
      give you (data is inherently point-in-time)
•EExactly what l
       tl    h t long-run means can be the subject of philosophical debate; I’ll
                                      b th      bj t f hil       hi l d b t
  define it as a customer’s average risk, given their characteristics, across the
  different economic conditions that might arise
    • This is not a lifetime estimate: eg their age/time on books doesn’t
      change
    • Which variables are considered to be cyclical can be a tricky decision to
      make (many behavioural variables eg credit card balance are probably
      correlated with the economy) y)
    • During bad economic times, the long-run PD will be below the spot, and
      vice-versa during good times
        • You don’t want to have to raise capital during a crisis



                                      Page 13
TTC approach

• Start with the spot estimate:

                              PD(x, e) = f(x, e)

• x = individual customer’s characteristics
• e = economic variables (constant for all customers at any point in time)
• Average over the possible values of e to get a TTC estimate
                   PD




                              Economic cycle




                                               Page 14
TTC approach

• This is complicated numerically, can be done in various ways eg Monte
  Carlo
• Use backcasting for simplicity: take historical values of e, substitute into
  prediction equation, average the results
    • As we are interested in means rather than quantiles, this shouldn’t
      affect accuracy much (other practical issues will have much more
      impact)
• R used to handle backcasting, combining multiple spot PDs into one output
  value




                                     Page 15
TTC calculation

• Input from spot model is a prediction equation, along with sample of
  historical economic data

spot_predict <‐ function(data)
{
  # code copied from SAS; with preprocessing, can be arbitrarily complex
  xb <‐ with(data, b0 + x1 * b1 + x2 * b2 + ... )
            (                                   )
  plogis(xb)
}

ttc_predict <‐ function(data, ecodata, from = "2000‐01‐01", to = "2010‐12‐01")
{
  dates <‐ seq(as.Date(from), as.Date(to), by = "months")
  evars <‐ names(ecodata)
  pd <‐ matrix(nrow(data), length(dates)) # not very space‐efficient!
  for(i in seq_along(dates))
  {
    data[evars] <‐ subset(ecodata, date == dates[i], evars)
    pd[, i] <‐ spot_predict(data)
  }
  apply(pd, 1, mean)
   pp y(p          )
}


                                                 Page 16
Binning/cohorting

• Raw TTC estimate is a combination of many spot PDs, each of which is from
  a logistic regression
      g        g
  → TTC estimate is a complicated function of customer attributes
• Need to simplify for communication, implementation purposes

• Turn into bins or cohorts based on customer attributes: estimate for each
  cohort is the average for customers within the cohort
• Take pragmatic approach to defining cohorts
    • Create tiers based on small selection of variables that will split out
      riskiest customers
    • Within each tier, create contingency table using attributes deemed most
      interesting/important to the business
    • Number of cohorts limited by need for simplicity/manageability, <1000
                                             simplicity/manageability
      desirable
    • Not a data-driven approach, although selection of variables informed by
      data exploration/analysis


                                    Page 17
SAS

                                      Model specification code                 Portfolio data         Economic data



R
                                            Empirical logits
      TTC averaging function                                                   Portfolio data         Economic data
                                             PIT predictor




                                                                                   PIT PD




       Cohorting functions             Cohorting specification                     TTC PD




      Empty cell imputation                Initial cohort table




                                           Final cohort table                                   Cohorted TTC PD



SAS

                     Initialisation code                          Cohorting code




                                                                   Page 18
Binning/cohorting

          Example from nameless portfolio:


                        Raw TTC PD                                                            Cohorted TTC PD
                        Distribution of ILS long-run PD                                        Distribution of cohorted ILS long-run PD
          0.4




                                                                                0.4
          0.3




                                                                                0.3
   sity




                                                                         sity
          0.2
            2
Dens




                                                                      Dens

                                                                                0.2
          0.1




                                                                                0.1
          0.0




                                                                                0.0




                0.01%    0.1%                1%           10%                         0.01%          0.1%                1%               10%

                                      PD                                                                         PD




                                                                Page 19
Binning input

 varlist <‐ list(
     low_doc2=list(name="low_doc",
                   breaks=c("N", "Y"),
                   midp=c("N", "Y"),
                    id   ("N" "Y")
                   na.val="N"),
     enq     =list(name="wrst_nbr_enq",
                   breaks=c(‐Inf, 0, 5, 15, Inf),
                   midp=c(0, 3, 10, 25),
                   na.val=0),
     lvr     =list(name="new lvr basel",
     lvr      list(name new_lvr_basel ,
                   breaks=c(‐Inf, 60, 70, 80, 90, Inf),
                   midp=c(50, 60, 70, 80, 95),
                   na.val=70),
     ...




                                                      low_doc wrst_nbr_enq new_lvr_basel ... tier1 tier2
                   by application of               1        N            0            50 ...     1     1
                  expand.grid() and                2        N            0            60 ...     1     2
                                                   3        N            0            70 ...     1     3
                       friends...
                       friends                     4        N            0            80 ...     1     4
                                                   5        N            0            95 ...     1     5
                                                   6        N            3            50 ...     1     6
                                                   7        N            3            60 ...     1     7
                                                   8        N            3            70 ...     1     8
                                                   9        N            3            80 ...     1     9
                                                   10       N            3            95 ...     1    10
                                                   10       N            3            95         1    10



                                                    Page 20
Binning output

if low_doc = ' ' then low_doc2 = 1;
else if low_doc = 'Y' then low_doc2 = 1;
else low_doc2 = 2;

if wrst_nbr_enq = . then enq = 0;
else if wrst_nbr_enq <= 0 then enq = 0;
else if wrst_nbr_enq <= 5 then enq = 3;
else if wrst_nbr_enq <= 15 then enq = 10;
else enq = 25;

if new_lvr_basel = . then lvr = 70;
else if new_lvr_basel <= 60 then lvr = 50;
else if new_lvr_basel <= 70 then lvr = 60;
else if new_lvr_basel <= 80 then lvr = 70;
else if new_lvr_basel <= 90 then lvr = 80;
else lvr = 95;
...




if lvr = 50 and enq = 0 and low_doc2 = 'N' then do; tier2 = 1; ttc_pd = ________; end;
else if lvr = 60 and enq = 0 and low_doc2 = 'N' then do; tier2 = 2; ttc_pd = ________; end;
else if lvr = 70 and enq = 0 and low_doc2 = 'N' then do; tier2 = 3; ttc_pd = ________; end;
 l   if l          d           d l   d      ' ' h    d    i              d               d
else if lvr = 80 and enq = 0 and low_doc2 = 'N' then do; tier2 = 4; ttc_pd = ________; end;
else if lvr = 95 and enq = 0 and low_doc2 = 'N' then do; tier2 = 5; ttc_pd = ________; end;
else if lvr = 50 and enq = 3 and low_doc2 = 'N' then do; tier2 = 6; ttc_pd = ________; end;
else if lvr = 60 and enq = 3 and low_doc2 = 'N' then do; tier2 = 7; ttc_pd = ________; end;
...




                                                    Page 21
Stress testing simulation

• Banks run stress tests on their loan portfolios, to see what a downturn
  would do to their financial health
• Mathematical framework is similar to the “Vasicek model”:
    • Represent the economy by a parameter X
    • Each loan has a transition matrix that is shifted based on X, determining
      its i k
      it risk grade in year t given its grade i year t - 1
                 d i           i    it     d in
        • Defaults if bottom grade reached
• Take a scenario/simulation-based approach: set X to a stressed value, run
  N times, take the average
          ,                g
    • Contrast to VaR: “average result for a stressed economy”, as opposed to
      “stressed result for an average economy”
• Example data: portfolio of 100,000 commercial loans along with current risk
  grade,
  grade split by subportfolio
• Simulation horizon: ~3 years




                                     Page 22
Application outline

• Front end in Excel (because the business world lives in Excel)
    • Calls SAS to setup datasets
       • Calls R to do the actual computations

• Previous version was an ad-hoc script written entirely in SAS, took ~4
  hours to run, often crashed due to lack of disk space
    • Series of DATA steps (disk-bound)
    • Transition matrices represented by unrolled if-then-else statements
      (25x25 matrix becomes 625 lines of code)
• Reduced to 2 minutes with R, 1 megabyte of code cut to 10k
    • No rocket science involved: simply due to using a better tool
    • Similar times achievable with PROC IML, of which more later




                                     Page 23
Application outline

• For each subportfolio and year, get the median result and store it
• Next year s simulation uses this year s median portfolio
       year’s                      year’s
• To avoid having to store multiple transited copies of the portfolio, we
  manipulate random seeds

for(p in 1:nPortfolios) # varies by project
f ( i 1 P tf li ) #          i   b     j t
{
  for(y in 1:nYears)    # usually 2‐5
  {
    seed <‐ .GlobalEnv$.Random.seed
                      $
    for(i in 1:nIters)  # around 1,000, but could be less
      result[i] <‐ summary(doTransit(portfolio[p, y], T[p, y]))

        med <‐ which(result == median(result))
        portfolio[p, y + 1] <‐ doTransit(portfolio[p, y], T[p, y], seed, med)
    }
}




                                                 Page 24
Data structures

• For business reasons, we want to split the simulation by subportfolio
• And also present results for each year separately
    → Naturally have 2-dimensional (matrix) structure for output: [i, j]th
     entry is the result for the ith subportfolio, jth year

• But desired output for each [i, j] might be a bunch of summary statistics,
  diagnostics, etc
    → Output needs to be a list

• Similarly, we have a separate input transition matrix for each subportfolio
  and year
    → Input should be a matrix of matrices




                                     Page 25
Data structures

R allows matrices whose elements are lists:

T <‐ matrix(list(), nPortfolios, nYears)
for(i in 1:nPortfolios) for(j in 1:nYears)
    T[[i, j]] <‐ getMatrix(i, j, ...)

M <‐ matrix(list(), nPortfolios, nYears)
M[[i, j]]$result <‐ doTransit(i, j, ...)
M[[i, j]]$sumstat <‐ summary(M[[i, j]]$result) 


• Better than the alternatives:
   • List of lists loses indexing ability
   • Separate matrices for each output stat is clumsy
   • Multi-way arrays conflate data and metadata




                                      Page 26
PROC IML: a gateway to R

• As of SAS 9.2, you can use IML to execute R code, and transfer datasets to
  and from R:

   PROC IML;
     call ExportDataSetToR('portfol', 'portfol');  /* creates a data frame */
     call ExportMatrixToR("&Rfuncs", 'rfuncs');
     call ExportMatrixToR("&Rscript", 'rscript');
     call ExportMatrixToR("&nIters", 'nIters');
     ...
     submit /R;
       source(rfuncs)
       source(rscript)
     endsubmit;
     call ImportDataSetFromR('result', 'result');
   QUIT;


• No messing around with import/export via CSV, transport files, etc
• Half the code in an earlier version was for import/export



                                           Page 27
IML: a side-rant

• IML lacks:
    • Logical vectors: everything has to be numeric or character
    • Support for zero-length vectors (you don’t realise how useful they are
      until they’re gone)
    • Unoriented vectors: everything is either a row or column vector
      (technically,
      (t h i ll everything i a matrix)
                         thi  is     t i )
• So something like x = x + y[z < 0]; fails in three ways

• IML also lacks anything like a dataset/data frame: everything is a matrix
    → It’s easier to transfer a SAS dataset to and from R, than IML

• Everything is a matrix: no lists, let alone lists of lists, or matrices of lists
   • Not even multi-way arrays

• Which puts the occasional online grouching about R into perspective



                                        Page 28
Other SAS/R interfaces

• SAS has a proprietary dataset format (or, many proprietary dataset
  formats) )
    • R’s foreign package includes read.ssd and read.xport for importing,
      and write.foreign(*, package="SAS") for exporting
    • Package Hmisc has sas.get
    • Package sas7bdat has an experimental reader for this format
                    7bd t
    • Revolution R can read SAS datasets
    • All have glitches, are not widely available, or not fully functional
        • First 2 also need SAS installed
• SAS 9.2 and IML make these issues moot
    • You just have to pay for it
    • Caveat: only works with R <= 2.11.1 (2.12 changed the locations of
      binaries)
        • SAS 9.3 will support R 2.12+




                                     Page 29
R and SAS rundown

• Advantages of R
   • Free! (base distribution, anyway)
   • Very powerful statistical programming environment: SAS takes 3
     languages to do what R does with 1
   • Flexible and extensible
   • Lots of features (if you can find them)
       • User-contributed packages are a blessing and a curse
   • Ability to handle large datasets is improving
• Advantages of SAS
   • Pervasive presence in large firms
       • “Nobody got fired for buying IBM SAS”
       • Compatibility with existing processes/metadata
              p      y              gp
       • Long-term support
   • Tremendous data processing/data warehousing capability
   • Lots of features (if you can afford them)
   • Sometimes cleaner than R, especially f d
                   l       h              ll for data manipulation
                                                            l

                                   Page 30
R and SAS rundown

Example: get weighted summary statistics by groups

   proc summary data = indat nway;
     class a b;
     var x y;
     weight w;
     output sum(w)=sumwt mean(x)=xmean mean(y)=ymean var(y)=yvar
       out = outdat;
   run;



   outdat <‐ local({
       res <‐ t(sapply(split(indat, indat[c("a", "b")]), function(subset) {
           c(xmean = weighted.mean(subset$x, subset$w),
             ymean = weighted.mean(subset$y, subset$w),
             yvar = cov.wt(subset["y"], subset$w)$cov)
       }))
       levs <‐ aggregate(w ~ a + b, data=indat, sum)
       cbind(levs, as.data.frame(res))                                    Thank god for plyr
   })


                                           Page 31
Challenges for deploying R

• Quality assurance, or perception thereof
    • Core is excellent, but much of R s attraction is in extensibility, ie
                                       R’s
      contributed packages
    • Can I be sure that the package I just downloaded does what it says? Is
      it doing more than it says?
    •B k
      Backward compatibility
                 d       tibilit
         • Telling people to use the latest version is not always helpful
• No single point of contact
    • Who do I yell at if things go wrong?
    • How can I be sure everyone is using the same version?
• Unix roots make package development clunky on Windows
    • Process is more fragile because it assumes Unix conventions
    • Why must I download a set of third-party tools to compile code?
    • Difficult to integrate with Visual C++/Visual Studio
• Interfacing with languages other than Fortran, C, C++ not (yet) integrated
  into core

                                    Page 32
Commercial R: an aside

• Many of these issues are fundamental in nature
• Third parties like Revolution Analytics can address them, without having to
  dilute R’s focus (I am not a Revo R user)
    • Other R commercialisations existed but seem to have disappeared
    • Anyone remember S-Plus?
    • Challenge is not to negate R’s drawcards
        • Low cost: important even for big companies
        • Community and ecosystem
            • Can I use Revo R with that package I got from CRAN?
            • If I use Revo R, can I/should I participate on R-Help,
              StackOverflow.com?
• Also why SAS, SPSS, etc can include interfaces to R without risking too
  much




                                     Page 33
Good problems to have

• Sign of R’s movement into the mainstream
• S-Plus now touts R compatibility, rather than R touting S compatibility
  S Plus
• Nobody cares that SHAZAM, GLIM, XLisp-Stat etc don’t support XML, C# or
  Java




                                  Page 34

More Related Content

PPT
The valuation of long-term securities
PPTX
PPT
US-Financial Crisis
PPTX
Ch8
PPT
Financial Management Ch 09
PPT
Business Statistics Chapter 3
PPT
Credit Risk Management Presentation
PPT
Financial Management Slides Ch 10
The valuation of long-term securities
US-Financial Crisis
Ch8
Financial Management Ch 09
Business Statistics Chapter 3
Credit Risk Management Presentation
Financial Management Slides Ch 10

What's hot (20)

PDF
The Value of Money - problems and solutions
DOCX
Business Strategy analysis Of Aarong
PPTX
Portfolio Optimisation
PPT
GITMAN Chapter 1
PPT
Value at risk
PPT
PPT
PPTX
Wealth Management and Behavioral Finance
PPTX
Chapter 7
DOCX
fiscal policy in Bangladesh
DOCX
Managerial Finance Chapter 10 solutions by Gitman 14 Edition
PPTX
Credit risk
PDF
Financial Statement Analysis and Valuation of Berger Paints Bangladesh Limited
PPT
Ch10
PPT
Chapter 7 - Stock Evaluation
PPT
6. operationalization of variables
PDF
Managerial Finance By Gitman Chapter 8 solutions
PPT
Principles of Managerial Finance
PDF
Credit Default Models
The Value of Money - problems and solutions
Business Strategy analysis Of Aarong
Portfolio Optimisation
GITMAN Chapter 1
Value at risk
Wealth Management and Behavioral Finance
Chapter 7
fiscal policy in Bangladesh
Managerial Finance Chapter 10 solutions by Gitman 14 Edition
Credit risk
Financial Statement Analysis and Valuation of Berger Paints Bangladesh Limited
Ch10
Chapter 7 - Stock Evaluation
6. operationalization of variables
Managerial Finance By Gitman Chapter 8 solutions
Principles of Managerial Finance
Credit Default Models
Ad

Viewers also liked (20)

PDF
R in finance: Introduction to R and Its Applications in Finance
PDF
Using R for Analyzing Loans, Portfolios and Risk: From Academic Theory to Fi...
PDF
Using R in Finance
PDF
Backtesting Trading Strategies with R
PDF
Business Finance KPI's - tutorial
PDF
How to use R in different professions: R In Finance (Speaker: Gabriel Foix, M...
PDF
Demystifying PD Terminologies
PPTX
Pairs Trading - Hedge Fund Strategies
PDF
ReStream: Accelerating Backtesting and Stream Replay with Serial-Equivalent P...
PPTX
Pairs trading using R
PPTX
The New Rules of Cross-Selling
PPTX
Cross sell - concept and related analytics
PDF
Introduction to pairtrading
PPTX
Machine Learning, Stock Market and Chaos
PDF
Analytics cross-selling-retail-banking
PPTX
Prezi v sway
PPTX
Pairs Trading from NYC Algorithmic Trading Meetup November '13
PPT
Cross Selling Opportunities In Banking Industry
PPTX
The Value of Open Source Communities
PPTX
DIY Quant Strategies on Quantopian
R in finance: Introduction to R and Its Applications in Finance
Using R for Analyzing Loans, Portfolios and Risk: From Academic Theory to Fi...
Using R in Finance
Backtesting Trading Strategies with R
Business Finance KPI's - tutorial
How to use R in different professions: R In Finance (Speaker: Gabriel Foix, M...
Demystifying PD Terminologies
Pairs Trading - Hedge Fund Strategies
ReStream: Accelerating Backtesting and Stream Replay with Serial-Equivalent P...
Pairs trading using R
The New Rules of Cross-Selling
Cross sell - concept and related analytics
Introduction to pairtrading
Machine Learning, Stock Market and Chaos
Analytics cross-selling-retail-banking
Prezi v sway
Pairs Trading from NYC Algorithmic Trading Meetup November '13
Cross Selling Opportunities In Banking Industry
The Value of Open Source Communities
DIY Quant Strategies on Quantopian
Ad

Similar to Successful Uses of R in Banking (20)

PPTX
Trust in banks
PDF
Historic interest rate charts
PDF
KW Office Locations by State Map - July 2010
PDF
T2 Partners Presentation On The Mortgage Crisis
PDF
T2 Partners Presentation On The Mortgage Crisis
PDF
2010 Member Profile: Indiana
PDF
Joanna Embry
PDF
Joanna Embry
PDF
Value PMS Motilal Oswal
PDF
Jason Mesut: Wicked Workshops
PDF
PDF
16193713 T2 Partners Presentation On The Mortgage Crisis
PDF
16193713 T2 Partners Presentation On The Mortgage Crisis
PDF
Alt picketude
PPTX
Questionnaire feedback slides
PDF
02 h.e. purnomo yusgiantoro
PPTX
Sony Nominations 2012
PPSX
Concept Design and Validation of LNG Powered Commuter Ferry
PDF
Global Internet M&A and Investment Analysis
PDF
Westfield Real Estate Market Report May 2012
Trust in banks
Historic interest rate charts
KW Office Locations by State Map - July 2010
T2 Partners Presentation On The Mortgage Crisis
T2 Partners Presentation On The Mortgage Crisis
2010 Member Profile: Indiana
Joanna Embry
Joanna Embry
Value PMS Motilal Oswal
Jason Mesut: Wicked Workshops
16193713 T2 Partners Presentation On The Mortgage Crisis
16193713 T2 Partners Presentation On The Mortgage Crisis
Alt picketude
Questionnaire feedback slides
02 h.e. purnomo yusgiantoro
Sony Nominations 2012
Concept Design and Validation of LNG Powered Commuter Ferry
Global Internet M&A and Investment Analysis
Westfield Real Estate Market Report May 2012

More from Revolution Analytics (20)

PPTX
Speeding up R with Parallel Programming in the Cloud
PPTX
Migrating Existing Open Source Machine Learning to Azure
PPTX
R in Minecraft
PPTX
The case for R for AI developers
PPTX
Speed up R with parallel programming in the Cloud
PPTX
The R Ecosystem
PPTX
R Then and Now
PPTX
Predicting Loan Delinquency at One Million Transactions per Second
PPTX
Reproducible Data Science with R
PPTX
The R Ecosystem
PPTX
R at Microsoft (useR! 2016)
PPTX
Building a scalable data science platform with R
PPTX
R at Microsoft
PPTX
The Business Economics and Opportunity of Open Source Data Science
PPTX
Taking R Analytics to SQL and the Cloud
PPTX
The Network structure of R packages on CRAN & BioConductor
PPTX
The network structure of cran 2015 07-02 final
PPTX
Simple Reproducibility with the checkpoint package
PPTX
R at Microsoft
PDF
Revolution R Enterprise 7.4 - Presentation by Bill Jacobs 11Jun15
Speeding up R with Parallel Programming in the Cloud
Migrating Existing Open Source Machine Learning to Azure
R in Minecraft
The case for R for AI developers
Speed up R with parallel programming in the Cloud
The R Ecosystem
R Then and Now
Predicting Loan Delinquency at One Million Transactions per Second
Reproducible Data Science with R
The R Ecosystem
R at Microsoft (useR! 2016)
Building a scalable data science platform with R
R at Microsoft
The Business Economics and Opportunity of Open Source Data Science
Taking R Analytics to SQL and the Cloud
The Network structure of R packages on CRAN & BioConductor
The network structure of cran 2015 07-02 final
Simple Reproducibility with the checkpoint package
R at Microsoft
Revolution R Enterprise 7.4 - Presentation by Bill Jacobs 11Jun15

Recently uploaded (20)

PPTX
Module5_Session1 (mlzrkfbbbbbbbbbbbz1).pptx
PDF
3CMT J.AFABLE Flexible-Learning ENTREPRENEURIAL MANAGEMENT.pdf
PDF
6a Transition Through Old Age in a Dynamic Retirement Distribution Model JFP ...
PPTX
social-studies-subject-for-high-school-globalization.pptx
PDF
The Right Social Media Strategy Can Transform Your Business
PDF
Pitch Deck.pdf .pdf all about finance in
PDF
7a Lifetime Expected Income Breakeven Comparison between SPIAs and Managed Po...
PDF
HCWM AND HAI FOR BHCM STUDENTS(1).Pdf and ptts
PPTX
Introduction to Customs (June 2025) v1.pptx
PPTX
Grp C.ppt presentation.pptx for Economics
PDF
3a The Dynamic Implications of Sequence Risk on a Distribution Portfolio JFP ...
PDF
Lecture1.pdf buss1040 uses economics introduction
PDF
THE EFFECT OF FOREIGN AID ON ECONOMIC GROWTH IN ETHIOPIA
PDF
Statistics for Management and Economics Keller 10th Edition by Gerald Keller ...
DOCX
BUSINESS PERFORMANCE SITUATION AND PERFORMANCE EVALUATION OF FELIX HOTEL IN H...
PDF
Chapter 9 IFRS Ed-Ed4_2020 Intermediate Accounting
PDF
1a In Search of the Numbers ssrn 1488130 Oct 2009.pdf
PPTX
FL INTRODUCTION TO AGRIBUSINESS CHAPTER 1
PDF
Blockchain Pesa Research by Samuel Mefane
PPTX
2. RBI.pptx202029291023i38039013i92292992
Module5_Session1 (mlzrkfbbbbbbbbbbbz1).pptx
3CMT J.AFABLE Flexible-Learning ENTREPRENEURIAL MANAGEMENT.pdf
6a Transition Through Old Age in a Dynamic Retirement Distribution Model JFP ...
social-studies-subject-for-high-school-globalization.pptx
The Right Social Media Strategy Can Transform Your Business
Pitch Deck.pdf .pdf all about finance in
7a Lifetime Expected Income Breakeven Comparison between SPIAs and Managed Po...
HCWM AND HAI FOR BHCM STUDENTS(1).Pdf and ptts
Introduction to Customs (June 2025) v1.pptx
Grp C.ppt presentation.pptx for Economics
3a The Dynamic Implications of Sequence Risk on a Distribution Portfolio JFP ...
Lecture1.pdf buss1040 uses economics introduction
THE EFFECT OF FOREIGN AID ON ECONOMIC GROWTH IN ETHIOPIA
Statistics for Management and Economics Keller 10th Edition by Gerald Keller ...
BUSINESS PERFORMANCE SITUATION AND PERFORMANCE EVALUATION OF FELIX HOTEL IN H...
Chapter 9 IFRS Ed-Ed4_2020 Intermediate Accounting
1a In Search of the Numbers ssrn 1488130 Oct 2009.pdf
FL INTRODUCTION TO AGRIBUSINESS CHAPTER 1
Blockchain Pesa Research by Samuel Mefane
2. RBI.pptx202029291023i38039013i92292992

Successful Uses of R in Banking

  • 1. Experiences with using R in credit risk Hong Ooi
  • 2. Introduction • A bit about myself: • Did an actuarial degree, found it terribly boring and switched to statistics after graduating • Worked for several years as a general insurance risk modeller • Switched to banking for a change of scenery (at the top of the boom) • Now with ANZ Bank in Melbourne, Australia • Contents: • LGD haircut modelling • Through-the-cycle calibration • Stress testing simulation app • SAS and R • Closing comments Page 2
  • 3. Mortgage haircut model • When a mortgage defaults, the bank can take possession of the property and sell it to recoup the loss1 p • We have some idea of the market value of the property • Actual sale price tends to be lower on average than the market value (the haircut)2 • If sale price > exposure at default, we don’t make a loss (excess is passed l i t d f lt d ’t k l ( i d on to customer); otherwise, we make a loss Expected loss = P(default) x EAD x P(possess) x exp.shortfall p ( ) (p ) p Notes: 1. For ANZ, <10% of defaults actually result in possession 2. Meaning of “haircut” depends on context; very different when talking about, say, US mortgages Page 3
  • 4. Sale price distribution p Valuation Expected shortfall Haircut Exposure at default $ Page 4
  • 5. Stat modelling • Modelling part is in finding parameters for the sale price distribution • Assumed distributional shape, eg Gaussian • Mean haircut relates average sale price to valuation • Spread (volatility) of sale price around haircut • Once model is found, calculating expected shortfall is just (complicated) arithmetic Page 5
  • 6. Valuation at origination Valuation at kerbside Valuation after possession 15 15 15 14 14 14 13 3 13 3 13 3 log sale price log sale price log sale price 12 12 12 11 11 11 10 10 10 9 9 9 11 12 13 14 15 10 11 12 13 14 15 10 11 12 13 14 15 log valuation log valuation log valuation Page 6
  • 7. Volatility Volatility of haircut (=sale price/valuation) appears to vary systematically: Property type SD(haircut)* A 11.6% B 9.3% 9 3% C 31.2% State/territory SD(haircut)* 1 NA 2 13.3% 3 7.7% 4 9.2% 5 15.6% 6 18.4% 7 14.8% * valued after possession Page 7
  • 8. Volatility modelling • Use regression to estimate haircut as a function of loan characteristics • Use log-linear structure for volatility to ensure +ve variance estimates log linear • Data is more heavy-tailed than Gaussian (even after deleting outliers), so extend to use t distribution • df acts as a shape parameter, controls how much influence outlying observations h b ti have • Constant volatility/Gaussian error model is ordinary linear regression • Varying-volatility models can be fit by generalised least squares, using gls Varying volatility in the nlme package; simpler and faster to directly maximise the Gaussian likelihood with optim/nlminb (latter will reproduce gls fit) • Directly maximising the likelihood easily extended to t case: just replace all *norm functions with *t norm t Page 8
  • 9. Normal model residuals 0.7 6 0.6 4 0.5 Sample quantiles 2 0.4 Density 0.3 0 0.2 -2 0.1 -4 0.0 -4 -2 0 2 4 6 -3 -2 -1 0 1 2 3 Std residual Theoretical normal quantiles t5-model residuals .4 0. 10 0.3 Sampl e quantiles 5 Density 0.2 0 0 0.1 -5 0.0 -5 0 5 10 -6 -4 -2 0 2 4 6 Std residual Theoretical t-quantiles Page 9
  • 10. Example impact Property type = C, state = 7, valuation $250k, EAD $240k Gaussian model Mean formula Volatility formula Expected shortfall ($) ~1 1 ~1 1 7,610 7 610 ~1 ~proptype 23,686 ~1 ~proptype + state 29,931 t5 model Mean formula Volatility formula Expected shortfall ($) ~1 ~1 4,493 ~1 ~proptype 10,190 ~1 ~proptype + state , 5,896 Page 10
  • 11. Notes on model behaviour • Why does using a heavy-tailed error distribution reduce the expected shortfall? • With normal distrib, volatility is overestimated → Likelihood of low sale price is also inflated • t distrib corrects this • Extreme tails of the t less important • At lower end, sale price cannot go below 0 • At higher end, sale price > EAD is gravy • This is not a monotonic relationship! At low enough thresholds thresholds, eventually heavier tail of the t will make itself felt • In most regression situations, assuming sufficient data, distributional assumptions (ie, normality, homoskedasticity) are not so critical: CLT comes into play • Here, they are important: changing the distributional assumptions can change expected shortfall by big amounts g g Page 11
  • 12. In SAS • SAS has PROC MIXED for modelling variances, but only allows one grouping variable and assumes a normal distribution • PROC NLIN does general nonlinear optimisation • Also possible in PROC IML • None of these are as flexible or powerful as R • The R modelling function returns an object, which can be used to generate predictions, compute summaries, etc • SAS 9.2 now has PROC PLM that does something similar, but requires the modelling proc to execute a STORE statement first • Only a few procs support this currently • If you’re fitting a custom model (like this one), you’re out of luck Page 12
  • 13. Through-the-cycle calibration • For capital purposes, we would like an estimate of default probability that doesn’t depend on the current state of the economy p y • This is called a through-the-cycle or long-run PD • Contrast with a point-in-time or spot PD, which is what most models will give you (data is inherently point-in-time) •EExactly what l tl h t long-run means can be the subject of philosophical debate; I’ll b th bj t f hil hi l d b t define it as a customer’s average risk, given their characteristics, across the different economic conditions that might arise • This is not a lifetime estimate: eg their age/time on books doesn’t change • Which variables are considered to be cyclical can be a tricky decision to make (many behavioural variables eg credit card balance are probably correlated with the economy) y) • During bad economic times, the long-run PD will be below the spot, and vice-versa during good times • You don’t want to have to raise capital during a crisis Page 13
  • 14. TTC approach • Start with the spot estimate: PD(x, e) = f(x, e) • x = individual customer’s characteristics • e = economic variables (constant for all customers at any point in time) • Average over the possible values of e to get a TTC estimate PD Economic cycle Page 14
  • 15. TTC approach • This is complicated numerically, can be done in various ways eg Monte Carlo • Use backcasting for simplicity: take historical values of e, substitute into prediction equation, average the results • As we are interested in means rather than quantiles, this shouldn’t affect accuracy much (other practical issues will have much more impact) • R used to handle backcasting, combining multiple spot PDs into one output value Page 15
  • 16. TTC calculation • Input from spot model is a prediction equation, along with sample of historical economic data spot_predict <‐ function(data) { # code copied from SAS; with preprocessing, can be arbitrarily complex xb <‐ with(data, b0 + x1 * b1 + x2 * b2 + ... ) ( ) plogis(xb) } ttc_predict <‐ function(data, ecodata, from = "2000‐01‐01", to = "2010‐12‐01") { dates <‐ seq(as.Date(from), as.Date(to), by = "months") evars <‐ names(ecodata) pd <‐ matrix(nrow(data), length(dates)) # not very space‐efficient! for(i in seq_along(dates)) { data[evars] <‐ subset(ecodata, date == dates[i], evars) pd[, i] <‐ spot_predict(data) } apply(pd, 1, mean) pp y(p ) } Page 16
  • 17. Binning/cohorting • Raw TTC estimate is a combination of many spot PDs, each of which is from a logistic regression g g → TTC estimate is a complicated function of customer attributes • Need to simplify for communication, implementation purposes • Turn into bins or cohorts based on customer attributes: estimate for each cohort is the average for customers within the cohort • Take pragmatic approach to defining cohorts • Create tiers based on small selection of variables that will split out riskiest customers • Within each tier, create contingency table using attributes deemed most interesting/important to the business • Number of cohorts limited by need for simplicity/manageability, <1000 simplicity/manageability desirable • Not a data-driven approach, although selection of variables informed by data exploration/analysis Page 17
  • 18. SAS Model specification code Portfolio data Economic data R Empirical logits TTC averaging function Portfolio data Economic data PIT predictor PIT PD Cohorting functions Cohorting specification TTC PD Empty cell imputation Initial cohort table Final cohort table Cohorted TTC PD SAS Initialisation code Cohorting code Page 18
  • 19. Binning/cohorting Example from nameless portfolio: Raw TTC PD Cohorted TTC PD Distribution of ILS long-run PD Distribution of cohorted ILS long-run PD 0.4 0.4 0.3 0.3 sity sity 0.2 2 Dens Dens 0.2 0.1 0.1 0.0 0.0 0.01% 0.1% 1% 10% 0.01% 0.1% 1% 10% PD PD Page 19
  • 20. Binning input varlist <‐ list( low_doc2=list(name="low_doc", breaks=c("N", "Y"), midp=c("N", "Y"), id ("N" "Y") na.val="N"), enq     =list(name="wrst_nbr_enq", breaks=c(‐Inf, 0, 5, 15, Inf), midp=c(0, 3, 10, 25), na.val=0), lvr     =list(name="new lvr basel", lvr list(name new_lvr_basel , breaks=c(‐Inf, 60, 70, 80, 90, Inf), midp=c(50, 60, 70, 80, 95), na.val=70), ... low_doc wrst_nbr_enq new_lvr_basel ... tier1 tier2 by application of 1        N            0            50 ...     1     1 expand.grid() and 2        N            0            60 ...     1     2 3        N            0            70 ...     1     3 friends... friends 4        N            0            80 ...     1     4 5        N            0            95 ...     1     5 6        N            3            50 ...     1     6 7        N            3            60 ...     1     7 8        N            3            70 ...     1     8 9        N            3            80 ...     1     9 10       N            3            95 ...     1    10 10 N 3 95 1 10 Page 20
  • 21. Binning output if low_doc = ' ' then low_doc2 = 1; else if low_doc = 'Y' then low_doc2 = 1; else low_doc2 = 2; if wrst_nbr_enq = . then enq = 0; else if wrst_nbr_enq <= 0 then enq = 0; else if wrst_nbr_enq <= 5 then enq = 3; else if wrst_nbr_enq <= 15 then enq = 10; else enq = 25; if new_lvr_basel = . then lvr = 70; else if new_lvr_basel <= 60 then lvr = 50; else if new_lvr_basel <= 70 then lvr = 60; else if new_lvr_basel <= 80 then lvr = 70; else if new_lvr_basel <= 90 then lvr = 80; else lvr = 95; ... if lvr = 50 and enq = 0 and low_doc2 = 'N' then do; tier2 = 1; ttc_pd = ________; end; else if lvr = 60 and enq = 0 and low_doc2 = 'N' then do; tier2 = 2; ttc_pd = ________; end; else if lvr = 70 and enq = 0 and low_doc2 = 'N' then do; tier2 = 3; ttc_pd = ________; end; l if l d d l d ' ' h d i d d else if lvr = 80 and enq = 0 and low_doc2 = 'N' then do; tier2 = 4; ttc_pd = ________; end; else if lvr = 95 and enq = 0 and low_doc2 = 'N' then do; tier2 = 5; ttc_pd = ________; end; else if lvr = 50 and enq = 3 and low_doc2 = 'N' then do; tier2 = 6; ttc_pd = ________; end; else if lvr = 60 and enq = 3 and low_doc2 = 'N' then do; tier2 = 7; ttc_pd = ________; end; ... Page 21
  • 22. Stress testing simulation • Banks run stress tests on their loan portfolios, to see what a downturn would do to their financial health • Mathematical framework is similar to the “Vasicek model”: • Represent the economy by a parameter X • Each loan has a transition matrix that is shifted based on X, determining its i k it risk grade in year t given its grade i year t - 1 d i i it d in • Defaults if bottom grade reached • Take a scenario/simulation-based approach: set X to a stressed value, run N times, take the average , g • Contrast to VaR: “average result for a stressed economy”, as opposed to “stressed result for an average economy” • Example data: portfolio of 100,000 commercial loans along with current risk grade, grade split by subportfolio • Simulation horizon: ~3 years Page 22
  • 23. Application outline • Front end in Excel (because the business world lives in Excel) • Calls SAS to setup datasets • Calls R to do the actual computations • Previous version was an ad-hoc script written entirely in SAS, took ~4 hours to run, often crashed due to lack of disk space • Series of DATA steps (disk-bound) • Transition matrices represented by unrolled if-then-else statements (25x25 matrix becomes 625 lines of code) • Reduced to 2 minutes with R, 1 megabyte of code cut to 10k • No rocket science involved: simply due to using a better tool • Similar times achievable with PROC IML, of which more later Page 23
  • 24. Application outline • For each subportfolio and year, get the median result and store it • Next year s simulation uses this year s median portfolio year’s year’s • To avoid having to store multiple transited copies of the portfolio, we manipulate random seeds for(p in 1:nPortfolios) # varies by project f ( i 1 P tf li ) # i b j t { for(y in 1:nYears)    # usually 2‐5 { seed <‐ .GlobalEnv$.Random.seed $ for(i in 1:nIters)  # around 1,000, but could be less result[i] <‐ summary(doTransit(portfolio[p, y], T[p, y])) med <‐ which(result == median(result)) portfolio[p, y + 1] <‐ doTransit(portfolio[p, y], T[p, y], seed, med) } } Page 24
  • 25. Data structures • For business reasons, we want to split the simulation by subportfolio • And also present results for each year separately → Naturally have 2-dimensional (matrix) structure for output: [i, j]th entry is the result for the ith subportfolio, jth year • But desired output for each [i, j] might be a bunch of summary statistics, diagnostics, etc → Output needs to be a list • Similarly, we have a separate input transition matrix for each subportfolio and year → Input should be a matrix of matrices Page 25
  • 26. Data structures R allows matrices whose elements are lists: T <‐ matrix(list(), nPortfolios, nYears) for(i in 1:nPortfolios) for(j in 1:nYears) T[[i, j]] <‐ getMatrix(i, j, ...) M <‐ matrix(list(), nPortfolios, nYears) M[[i, j]]$result <‐ doTransit(i, j, ...) M[[i, j]]$sumstat <‐ summary(M[[i, j]]$result)  • Better than the alternatives: • List of lists loses indexing ability • Separate matrices for each output stat is clumsy • Multi-way arrays conflate data and metadata Page 26
  • 27. PROC IML: a gateway to R • As of SAS 9.2, you can use IML to execute R code, and transfer datasets to and from R: PROC IML; call ExportDataSetToR('portfol', 'portfol');  /* creates a data frame */ call ExportMatrixToR("&Rfuncs", 'rfuncs'); call ExportMatrixToR("&Rscript", 'rscript'); call ExportMatrixToR("&nIters", 'nIters'); ... submit /R; source(rfuncs) source(rscript) endsubmit; call ImportDataSetFromR('result', 'result'); QUIT; • No messing around with import/export via CSV, transport files, etc • Half the code in an earlier version was for import/export Page 27
  • 28. IML: a side-rant • IML lacks: • Logical vectors: everything has to be numeric or character • Support for zero-length vectors (you don’t realise how useful they are until they’re gone) • Unoriented vectors: everything is either a row or column vector (technically, (t h i ll everything i a matrix) thi is t i ) • So something like x = x + y[z < 0]; fails in three ways • IML also lacks anything like a dataset/data frame: everything is a matrix → It’s easier to transfer a SAS dataset to and from R, than IML • Everything is a matrix: no lists, let alone lists of lists, or matrices of lists • Not even multi-way arrays • Which puts the occasional online grouching about R into perspective Page 28
  • 29. Other SAS/R interfaces • SAS has a proprietary dataset format (or, many proprietary dataset formats) ) • R’s foreign package includes read.ssd and read.xport for importing, and write.foreign(*, package="SAS") for exporting • Package Hmisc has sas.get • Package sas7bdat has an experimental reader for this format 7bd t • Revolution R can read SAS datasets • All have glitches, are not widely available, or not fully functional • First 2 also need SAS installed • SAS 9.2 and IML make these issues moot • You just have to pay for it • Caveat: only works with R <= 2.11.1 (2.12 changed the locations of binaries) • SAS 9.3 will support R 2.12+ Page 29
  • 30. R and SAS rundown • Advantages of R • Free! (base distribution, anyway) • Very powerful statistical programming environment: SAS takes 3 languages to do what R does with 1 • Flexible and extensible • Lots of features (if you can find them) • User-contributed packages are a blessing and a curse • Ability to handle large datasets is improving • Advantages of SAS • Pervasive presence in large firms • “Nobody got fired for buying IBM SAS” • Compatibility with existing processes/metadata p y gp • Long-term support • Tremendous data processing/data warehousing capability • Lots of features (if you can afford them) • Sometimes cleaner than R, especially f d l h ll for data manipulation l Page 30
  • 31. R and SAS rundown Example: get weighted summary statistics by groups proc summary data = indat nway; class a b; var x y; weight w; output sum(w)=sumwt mean(x)=xmean mean(y)=ymean var(y)=yvar out = outdat; run; outdat <‐ local({ res <‐ t(sapply(split(indat, indat[c("a", "b")]), function(subset) { c(xmean = weighted.mean(subset$x, subset$w), ymean = weighted.mean(subset$y, subset$w), yvar = cov.wt(subset["y"], subset$w)$cov) })) levs <‐ aggregate(w ~ a + b, data=indat, sum) cbind(levs, as.data.frame(res)) Thank god for plyr }) Page 31
  • 32. Challenges for deploying R • Quality assurance, or perception thereof • Core is excellent, but much of R s attraction is in extensibility, ie R’s contributed packages • Can I be sure that the package I just downloaded does what it says? Is it doing more than it says? •B k Backward compatibility d tibilit • Telling people to use the latest version is not always helpful • No single point of contact • Who do I yell at if things go wrong? • How can I be sure everyone is using the same version? • Unix roots make package development clunky on Windows • Process is more fragile because it assumes Unix conventions • Why must I download a set of third-party tools to compile code? • Difficult to integrate with Visual C++/Visual Studio • Interfacing with languages other than Fortran, C, C++ not (yet) integrated into core Page 32
  • 33. Commercial R: an aside • Many of these issues are fundamental in nature • Third parties like Revolution Analytics can address them, without having to dilute R’s focus (I am not a Revo R user) • Other R commercialisations existed but seem to have disappeared • Anyone remember S-Plus? • Challenge is not to negate R’s drawcards • Low cost: important even for big companies • Community and ecosystem • Can I use Revo R with that package I got from CRAN? • If I use Revo R, can I/should I participate on R-Help, StackOverflow.com? • Also why SAS, SPSS, etc can include interfaces to R without risking too much Page 33
  • 34. Good problems to have • Sign of R’s movement into the mainstream • S-Plus now touts R compatibility, rather than R touting S compatibility S Plus • Nobody cares that SHAZAM, GLIM, XLisp-Stat etc don’t support XML, C# or Java Page 34