Synthetic Data, never run out of OOS data again.

[MeiHua]

Here is another simple exercise anyone can use. its great because it generates unlimited amounts of data for you to back test over, practice on, evaluate a system, or just look at. The reason why this method which is a spin off of what Tuschar Chande does is because it uses the entire range of data you set instead of random sampling. With random sampling you could resample moves, here each move is only used once. Also this does a lot of the same work as my data mining system. It maintains the same distribution of movement as the market actually had over that period, even though it is using a randomize that is normal in distribution. So if you ever wanted more practice on a single pair but have used all the historical data your provider gave you, you can generate more that will maintain that specifics markets personality.

Here is a picture of some real vs synthetic data I generated its the 1st 100 bars of a 15 minute chart.


You can see that in the synthetic price patter you still have, dojis, engulfing bars, quiet times as well as large range bars. This is a true test because if your system can still come out positive in this environment it is able to pick out something that is NON-RANDOM and there is no chance for you to curve fit, as you can recycle this data a million times and arrive at a million different price structures.

here a pic of the spreadsheet you can make it yourself its easy enough.




Instructions:
1) export and cut and paste the time and price data
2) calculate the differences between the OHLC of the bar to the C of the previous bar
3) copy values and add rand() function
4) sort values by rand() column
5) use the close of the 1st candle of real data as reference, calculate new bars based on the scrambled movements
6) save and format to be imported in to the charting package.
7) enjoy

ps: I got the spreadsheet layout from a different forum. I just thought I would share it, but i cant link it.

[ClarkFX]

You're going to have a hard time getting accurate testing results from randomly generated data because it does not completely reflect true market behaviours. Rather than thinking the market is completely random, I'd like to think the market is pseudo-random.

[MeiHua]

You're going to have a hard time getting accurate testing results from randomly generated data because it does not completely reflect true market behaviours. Rather than thinking the market is completely random, I'd like to think the market is pseudo-random.

I completely agree with you. Because i isolated the volatility, this is not random, it remains as fat tailed as the original distribution. The only difference is the order in which these events occur, so your edge should still remain assuming that your exploiting some thing that exists that is non random, that occurs in the market no matter what order the candles are given.

[Jezzode]

Excuse me for being old fashioned, but it's far easier to use real historical data. Nothing better than the real thing! :P
30 pairs, 10 years on each...who would need any more data!

[MeiHua]

That point is taken. 7,5 years In sample, 2.5 out of sample is usually my standard deal. Using uncorrelated or correlated pairs is also useful. But I could always use more with the same dynamics as the market I am trying to trade it on. In general probably no ones going to use this. But when i make something i need 30+ System years of solid data before i make a decision.

[ClarkFX]

That point is taken. 7,5 years In sample, 2.5 out of sample is usually my standard deal. Using uncorrelated or correlated pairs is also useful. But I could always use more with the same dynamics as the market I am trying to trade it on. In general probably no ones going to use this. But when i make something i need 30+ System years of solid data before i make a decision.

There's a few recent articles stating that the average life of a positive return algorithm is around 3 months. Trying to create a robust system across multiple markets and such large samples is tough. Also, with regime shifts I find it redundant to test on such large samples.

[MeiHua]

There's a few recent articles stating that the average life of a positive return algorithm is around 3 months. Trying to create a robust system across multiple markets and such large samples is tough. Also, with regime shifts I find it redundant to test on such large samples.

Do your systems not incorporate a regime shift mechanism? Or do you just turn some on and some off? Assuming those regimes return in some point in the futures, as there is some cyclical/seasonal effect? I would love to check those articles if you could link em. For me its about robustness and portfolio diversification. To be honest I am pretty green behind the ears here since i have never traded live with an algo. So any advice would be good. Just taking what i see out there, and what makes sense.