1 #+date: <2017-02-17 20:07:10 +0800>
2 #+filetags: :finance:fintech:analysis:forecasting:Markov:
3 #+title: Loans and semi-Markov chains
5 Using Markov chains' transition matrices to model the movement of loans from being opened (in a state of
6 "Current") to getting closed can misinform the user at times.
8 To illustrate the challenge, the graph below plots the evolution, from
9 the original state to the final state, of a group of loans over 6
12 #+CAPTION: Actual vs predicted loan vintage performance.
13 #+ATTR_HTML: :width 400 :class img-fluid :alt Actual vs predicted loan vintage performance.
14 [[file:../assets/rollRateBeware.png]]
16 The solid lines are the result of applying an average transition matrix
17 6 times (the model's predicted outcome). The dashed lines are the actual
18 observed results for a set of loans.
20 As can be seen, the model does not do a very good job at predicting the
21 accounts that will end up in state "Closed" in each period. They end up
22 in a different state between Current and Closed (i.e. overdue) at a
23 higher than expected rate. Why is that?
25 The prediction was built using an average of the transition matrix of a
26 number of consecutive period statetables for a book of loans. That book
27 was not homogenic though. Most obviously, the "Current" accounts were
28 not of the same vintage - some had been in that state for a number of
29 periods before already. The observed set of loans all originated in the
30 same period. Other differences can be related to client demographics,
31 loan characteristics or macro-economic circumstances.
33 Applying a transition matrix based on a group of loans of various
34 vintages to a group of loans that all were new entrants in the book
35 violates the often implied Markov chain assumption of time-homogenity.
37 What that assumption says is that the future state is independent of the
40 Loans typically have a varying chance of becoming delinquent in function
41 of how long they have been open already.
43 Multi-order Markov chains are those that depend on a number (the order)
44 of states in the past. The question becomes - what order is the Markov
45 chain? Otherwise put, how many previous periods need to be taken into
46 account to be able to accurately estimate the next period's statetable?
47 Controlling for the other differences suggested above, if found to be
48 material, may be important as well.