I think the tool will depend on your use case. Are you testing a large portfolio of instruments? How much historical data are you using and how many sources are you having to pull the data from?

Hi Saurabh, please have a look at our website and let me know if interested.

http://systemathics.com/backtest.htm

Best regards,

Salah

Hello, which problem do you see by using R ?

Hi Rob, I am trying to backtest my algo on some 80k data points and each point has some 4-5 if then else loops. I take 40 mins and 1 hr on VB and R respectively. If there any other tool or language i can use to speed up my calc.

Dear Salah,I am more keen to do this myself and play with the variables at this stage of development and want to outsource the work.

regards

Hi Saurabh, it is curious that a backtesting on 80K*5 data points takes about one hour.

Did you try to isolate each step of the backtesting and see who does what ?

My recommendation is to split the process into 3 different steps :

1- normalize market data

2- simulation step (signals and portfolio building)

3- post-trade analysis step (perf, optimization ...)

Also, maybe you have to multi-thread some pieces ...

Hope this is helpful.

Best regards,

Salah

We're supporting R.NET inside QuantConnect soon. You should check us out. It would be the best of both worlds - local and cloud backtesting...

Try quantopian.com for backtesting. If you have ur own data for backtest use zipline package in python for backtest (Given complexity you mentioned with data points i m sure it would take at max few mins for backtesting)

Have you inspected GPU computing in R as a way to speed up your backtesting? http://www.r-bloggers.com/?s=GPU

I know you can do better. Just for fun, I tested 400 days each for 400 equity's for a simple strategy. 12 sec. C++ for the program, SQL for the data storage/retreval.

First, question that I would ask is : What kind of code executes within each "If Then" condition?.

The reason for that question is that a simple loop with 4-5 simple conditions (i.e. add 1 to a counter) should take less then 1 second on a modern PC. So the problem is not in the loop but rather in the code that is executed within the condition statements. Frequently, I have found when debugging such code that the programmer reloaded the data within each iteration within the loop. That is not to say what is happening here, but there maybe something else like that happening within those condition statements.

Right Sam, I am applying counter on each loop, i should rather use one counter for all the loops.. this would save me some time..

OP, if you show some core code of the if-then-self and loops(?) which you think is the most time consuming, we may help to see how to optimize your code. 80K data takes one hour to process is not normal, unless you are doing very time consuming job.

I choose C++ to develop my backtesting/trade system because I can gain the best performance from C++.

The only problem is time Manuele, perhaps my code is not very efficient. I will try to refine it.

Thanks Salah, I will try and split n process to check the problem with my code.

Good to know about the GPU computing Aatos, i will read up on the same. Thanks very much..

Thanks Avinash for suggesting Python.

Right Colin, C++ is indeed a faster option available, i will switch to that if R doesn't serve the purpose..

Just of out curiosity wondering what is it that you folks are backtesting? Exposure or MTM? Have you folks tried python using large numpy arrays (assuming you have robust data store for storage & retrieval of large time series based data)? Thanks

@ Saurabh : I use R on multiples cores and servers. If you can't split your task on multiple cores but you must use only one core try in C but it takes time to code it. R is faster to setup at an higher level of research.

I usually run R for hours of computational time on multiple tests. To value the test you need time, so it is not clear to me which is the problem for 45 min ? Do you have to run it multiple times during the same day ?

I agree with some of the folks here, the amount of time necessary for 80K data points x 4-5 conditions taking 1h for R and much less in VBA just don't seem right.

I would profile the R code and determine where the bottleneck is first. Then use R packages like Parallel/Multicore/Snow to parallelize your code. Quite often you'll find significant trimming down in execution time to only a small fraction of your original time.

If you need to run your code in real time, then it's best to translate it to C/C++. If it's acceptable to run your code every few minutes, then spend the effort to vectorize your algo and parallelize them. You are not talking about a lot of data after all. I wouldn't be surprised if your operation can be trimmed down to less than 1 or 2 minutes. For me, trimming it down to a minimum is lot of fun.

Good luck and have fun while doing it.

Optimal Jforex platform ( java )

You shouldn't be using loops in R if you can get away with it. Looping is horrendously inefficient in R. R is much much better at vector operations then looping. So if you can turn your loop into a vector operation the speed gains will be enormous. Trade of will be that your're using a lot more memory.

Example:

Increment the values of a vector x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] by 1

Solution 1: Use an if then else loop (slow)

Solution 2: Create a vector y = [1,1,1,1,1,1,1,1,1,1]. x = x + y. Much faster. Obviously also more memory intensive.