Quote from gmst:
Agree with most of what you say - just want to bring to your attention though that some funds might be trading with a multi-month hold periods. If trade hold period is in months, 72 trades is a very reasonable and respectable number to have in a backtest. E.g. consider going long when MA50 crosses MA200.
I haven't read Vishnu's book - so I might be off here. Have no idea of his strategies. My gut says - Vishnu didn't work as hard as we pikers do, because I have strategies which backtest for 8 yrs and have been consistently profitable for 2 yrs. Backtest had 400 trades or so, forward test 100 trades or so
Edit: Also consider if hypothetically a strategy tests for a date effect or something like that, over the last 10 yrs, you will get only 10*12 = 120 data points. So, Vishnu is correct in saying that depending on model, even 20 trades can be significant. What if you test only for Fed days - surely the whole population is limited - so your sample set can't be too big.
I see your points but as his trading was done on the daily scale, 72 trades is not significant.
I haven't traded on a monthly or weekly scale but he mentioned 20 trades being significant - I'd completely disagree about this. The question becomes, would you trade real money with confidence based on that? I certainly wouldn't.
I've done a lot of research regarding FED days and looked into dates in every form I could think of, I discovered something worthwhile and had at least 50 data points but never implemented it for the above mentioned reasons. When dealing with rare events, you need a large sample. You can curve fit and find the "perfect" market lows during crashes but it's still a curve fit.
The problem with his statement is that he cannot say his methods were valid, YET the edges stopped working after discovered. There is a fallacy here.
I've tested the strategies provided in the book as they're all publicly available, there was quite a lot of interest surrounding it when it came out but that quickly faded.
