Through the crisis : UK SMEs performance during the ‘credit crunch’
View/ Open
Ma2017.docx (3.860Mb)
Date
10/07/2017Item status
Restricted AccessAuthor
Ma, Meng
Metadata
Abstract
The influence of ‘credit crunch’ on Small and Medium sized Enterprises (SMEs) has
been of concern to the government, regulators, banks, the enterprises and the public.
Using a large dataset of UK SMEs’ records covering the early period of the ‘credit
crunch’, the influence of the ‘credit crunch’ on SMEs have been studied. It uses cross-sectional
method, panel data models and GAM to provide a detailed examination of
SMEs performance. Both newly established and matured SMEs, segmented by age,
are considered separately. The data contains 79 variables which covered obligors’
general condition, financial information, directors’ portfolio and other relevant credit
histories.
The ‘credit crunch’ is a typical ‘black swan’ phenomenon. As such there is a need to
examine whether the stepwise logistic model, the industries prime modelling tool,
could deal with the sudden change in SMEs credit risk. Whilst it may be capable of
modelling the situation alternatives models may be more appropriate. It provides a
benchmark for comparison to other models and shows how well the industry’s
standard model performs. Given cross-sectional models only provide aggregative
level single time period analysis, panel models are used to study SMEs performance
through the crisis period. To overcome the pro-cyclic feature of logistic model,
macroeconomic variables were added to panel data model. This allows examination of
how economic conditions influence SMEs during ‘credit crunch’. The use of panel
data model leads to a discussion of fixed and random effects estimation and the use of
explanatory macroeconomic variables. The panel data model provides a detailed
analyse of SMEs’ behaviour during the crisis period.
Under parametric models, especially logistic regression, data is usually transformed to
allow for the non-linear correlation between independent variable and dependent
variable. However, this brings difficulty in understanding influence of each
independent variable’s marginal effects. Another way of dealing with this is to add
non-parametric effects. In this study, Generalized Additive Models (GAM) allows for
non-parametric effects. A natural extension of logistic regression is a GAM model
with logistic link function. In order to use the data in their original state an alternative
method of processing missing values is proposed, which avoids data transformation,
such as the use of weights of evidence (WoE). GAM with original data could derive a
direct marginal trend and plot how explanatory variables influence SMEs’ ‘bad’ rate.
Significant non-parametric effects are found for both ‘start-ups’ and ‘non-start-ups’.
Using GAM models results in higher prediction accuracy and improves model
transparency by deriving explanatory variables’ marginal effects.