No. of Recommendations: 17
In 1999-2000 particularly there was a great amount of MI research taking place here despite the availability of data was exceedingly lower. Some of the authors such as Kuperman had been going through text books in the library and typing in the data, and Valueline was really treasured, even most of the interesting fields at that time only went a few years back.
With such wonderful backtesting tools now, we have a lot of opportunities to create truly novel MI approaches.
Among other MI subject am interested in backtesting broad market returns with meaningful (and novel) psychological sentiment indexes.
Kudos to authors who have been posting real world results from their actual MI activity - I think this is a bedrock of MI, given how easily it is to fall into the darkness of data mining (whilst performing backrests) without realising it.
For example, you modify the parameters of a screen to improve the backtested result, or you merely skip many screens and until a 'good' one is found (high backtested CAGR), in *both* cases one is falling for selection bias. You can mitigate this by looking at robustness when repeating with stock selections further down the final ranking, as well as adjusting parameters but keeping the same 'economic sense' of the screen, and making sure the results don't fall apart.
If there are 100 screens that select stocks 100% randomly, and you backtested five and chose the best, then years later you select another five screens, and replaced the former best with the new best, one can start to think we are not dealing with random screens. Yet as the screens are random stock selectors, the screen with the lowest CAGR will have the same future prospects as the screen with the highest CAGR.
- Manlobbi
