Key Performance Indicators (KPIs)

[algorithmbetting]

For those of you monitoring performance of systems and/or backtesting those models I wondered what KPIs you consider and what importance you place on each of these. Personally, I consider performance by month and in particlaur ROI/yield and standard deviation of ROI/yield by month. From this i monitor one single figure, coefficient of variation = (stdev of avg monthly ROI)/(avg monthly ROI). Other variables such as number of bets, nominal fund returns, strike rate obviously all have their place. I raise this as I would rather just use one figure in order to rank systems but find myself flitting between all of them. I guess it depends on your level of risk tolerance? Perhaps someone has developed a ranking system for system outputs!!:rollin

[Sgt.Sunshine]

Re: Key Performance Indicators (KPIs) Very interesting question. I guess there are different answers for different systems and sports, and it will also vary depending on whether they are long odds e.g. 33/1 or shorties as for football. I pretty much go on yield to unit stakes to assess a model's performance (assuming a large sample size = several years), and manually look at how it performs (or performed in back testing) month to month, or more realistically year to year. As a long term strategy there is no point in following a system that has not provided +ve yields EVERY year IMO. This doesn't of course guarantee future performance and system performance should be monitored. Bank balance is always a good indicator :lol. In terms of ranking or comparing systems, I thought of a way some time ago but never really tried it out. A good system is a combination of unit yield AND consistency, i.e. a system may yield 10% but it might have a negative yield for the most part, with a good run at the end. To measure consistency I was thinking along the lines of a regression, and maybe using r-sq, but I don't know how comparable that would be so I thought of using a Spearman Rank of profit to unit stakes against time. SR has a max value of 1, so you could maybe multiply the unit yield * SR to get an idea of how good a system is?