What will we mean by financial crisis? What classic methods are there to predict such crises? How can machine learning algorithms help anticipate these?
Financial crises can take many forms: they vary from government insolvencies to bank runs and currency crises. What all of those episodes have in common is that an internal vulnerability worsens over time and, following an associated trigger, triggers a financial crisis.
It may be difficult to find out the precise trigger, so the event of internal vulnerabilities have to be monitored. What exactly are these internal vulnerabilities? Statistically speaking, they’re the explanatory variables in crisis models. They often served as response variables in historical crisis episodes.
Although this is a component of the classic approach to modeling financial crises, it shouldn’t be the one solution to model financial risks.
In the classic crisis model, the usual method is to estimate the probability of a financial crisis using logistic regressions. Explanatory variables are linked to the response variable using a nonlinear link function. The dependent variable is 0 for no crisis and 1 for crisis. This approach is determined by the definition of the financial crisis. The past variables are modeled using the utmost likelihood method by various the exposure of the explanatory variable to the response variable. In terms of machine learning, this can be a supervised learning technique or logistic regression with a hidden layer. It can also be called a shallow neural network.
Other crisis modeling methods include determining default or crisis probabilities from market prices. For example, an implicit probability of default may be calculated from credit default swaps (CDS). Of course, that is fundamentally different from each logistic regression described above and the appliance of machine learning algorithms described below.
So what can machine learning algorithms do to enhance estimates of the likelihood of a financial crisis? First, unsupervised learning differs from supervised learning in that there is no such thing as a response variable. Clustering is a way value highlighting. The goal of clustering is to group data points in a meaningful way. These groups of knowledge are assigned a middle of mass to find out the structure throughout the data sets. Clustering may be applied to each the dependent and independent variables. For example, as an alternative of using a set threshold to find out a currency crisis, we are able to divide currency returns into different clusters and derive meaningful meaning from each cluster.
Machine learning algorithms can add significant value in this fashion. While clustering is only one example of the ability of coding, these algorithms have a variety of other useful applications
While machine learning is just an umbrella term for a lot of useful algorithms, whether the machine actually learns is a totally different query.
However, dividing the time series right into a training and testing set continues to be considered one of the most important weaknesses of machine learning. How do you establish the division? Often the choice is bigoted.
Whatever these shortcomings could also be, they hardly detract from the numerous advantages that machine learning can bring. In fact, now’s the time to speculate in these skills.
If you enjoyed this post, remember to subscribe
Photo credit: ©Getty Images/noLimit46
