I used financial ratios to predict financial distress using machine learning algorithms. The data consist of 2002-2015. Economic shocks effect the ratios and may affect the generalizability of the prediction model. Can using industry adjusted ratios solve this problem and control the economic shocks to develop more robust prediction model?
Most likely not, There’s an important relationship here between economic shocks and company level financial ratios you want to capture. Exogenous shocks may affect company top lines, input costs and by implication margins and the bottom line. This in turn will affect the companys ability to service debt. As such you may want to include a variable that indicates the prevailing economic regime as well. An example would be say the infrastruture sector which is usually highly leveraged and sensitive to changes in the discount rate, in an economic regime where the real interest rate is rising listed infrastructure could re-rate heavily. This also plays to your point about industry adjustments, different industries will have different sensitivities to economic shocks but they are still governed by the economic conditions of the economy they operate in. I’ve used shocks and regimes here separately deliberately but shocks often lead to regime changes and changes in ratios (a increasingly levered balance sheet in an inflationary growth regime might not be as good a predictor as say in a growth slowdown envrionment ) it is also extremely difficult to predict shocks, as the name suggests. It might also be worth adding an industry/sector variable.
At least they can help, but it depends of the way you do the adjustment. You must choose how to do it differences to the mean or the median ? windsorizing ? Additionally you can to do a panel data analysis with a dummy variable for economic shocks.
Well, it depends on the ratio and what type of economic shocks you try to control. I don't believe that there is (or will be) a general model for economic shocks. Different shocks could be foreseen with different financial indicators. I'm not saying that we know which ratios can be associated with which shocks, but only that such (most likely very complex) relationship has to exist.
As Yifan Wang suggested, predicting shocks is a very difficult and sophisticated task. However, it's not like humans have not been able to predict any shocks. A (very bad) example would be predicting natural disaster. It is worth noting from this bad example that the information required for a prediction might not be obvious, e.g. small changes in the temperature in west africa might lead to major hurricanes in US east coasts and the caribics.
Limiting your (I assume unsupervised) machine learning algorithm for industry ratios might not be smart for your goal. Assuming (past) industry ratios are significant predictors for (near-future) economic shocks, then it is clear that linear adjustment will not lead you to anywhere (when using neural networks). I'm not sure whether the same can be asserted for non-linear adjustments and non-linear machine learning techniques. I guess, the answer will be different for different techniques.
- Badges
- Science topic