Modelowanie w ryzyku

ZASŁYSZANE:

When it comes to modelling, I always thought it was a bit funny to see the world of the academic models compete with world of the operational models. One being correct and advanced but nearly impossible to implement in the business due to the facts of real life without needing to make assumptions and constrains so big that the model lost all its beauty and “correctness” and becoming in-transparent huge monsters (although giving the result with at least five digits). And the other ones being insufficient or becoming engineered into adhoc-hell sometimes sufficient for giving predictions (one scenario forecast) but worthless for risk and other times worthless on both. RiskSim (switching regime copula hybrid model) and FCSim (forward curve simulation with steered dependencies) as examples.

+ & - 10% isn’t that bad when it comes to transparency but it then the result demands some understanding from the interpreter. Or my other favourite that was the so-called profit curve where you made some assumptions and regressions (often linear) on the coming profit e.g. monthly profit with the forward spread going into delivery as a variable. You can then calculate the risk around the curve (discrete in sections of the spread is most likely due to the number of data points). Secondly you can now simulate the forward curve in order to get the risk from today to the delivery month. In order to get more sophisticated one could always increase the number of variable in the regression adding for example the hydrological balance, average temperature. This will then risk to add the complexity in the FC-simulation. And as always one should do some significance test on all variables used. However one of the beauty of this approach is that is quite transparent and understandable in interface of the two models. And also it has the advantage that one could use historical distribution for the profit curve and/or model adjusted scenarios if one do not expect the history to repeat itself (in this aspect). Also one could stress to the extreme in the FC scenarios and still have some accuracy if the profit curve is within the historical boundaries.