I'm new too. I have one algorithm that is fairly intensive (aka: inefficient) , as it grows it gets worse and I can see the slowdown over time.

How long is your backtesting? In hours?

Backtesting time isn't great. I can't really backtest any descent machine learning-based hypotheses. Anything more than a few features with a training window of more than 50 days or so makes the backtest unapproachable. In particular, any data transformations from original feature space to a transformed space before classification are not doable. Today, I tried one on just a few months of calendar time and after ~6 hours it was still only ~3% done.

Mine's pretty simple compared to Ian, but over the course of 10yrs backtesting I can typically get through about 4-5 years in a couple hours then years 6-10 take overnight.

Thank you all for your inputs.
Looks like I'm not alone...mine even got timed out if the dataframe I created is too big.
I tried to eliminate loops if possible. I think we can work harder to optimize the code but there is always limit.
Quantopian has to support all the coders so no way we can get super fast computing resource.

It's taking me about 5 minutes to get through a year of backtesting. The platform works fine for what I'm doing, but losing the internet connection even for a few seconds will stop the backtest, which is the biggest issue. Pipeline-based algorithms seem a lot slower than older algorithms but the pipeline is much more flexible and the older algorithms may stop working at some point. I haven't tried to use machine learning yet.

Thanks for everyone's feedback on this thread. It pains me to read any report that our backtester is too slow. We are actively working on speeding up some of the data sources that are bottlenecks to many backtests.

We have a bunch of best practices here: https://www.quantopian.com/quantopian2/best_practices. They boil down to accessing data only when you need it (for example, accessing context.portfolio requires recomputing your portfolio, including fetching the last known prices of any assets you hold, so it's relatively expensive to do every minute of a backtest).

If would be great if you guys could send me a version of your algorithm with any proprietary ideas removed, or some other example algorithm that lets me see the bottleneck you are experiencing. We can then help identify any optimizations that can be made, or otherwise learn about a new bottleneck in our system. My email address is my first name at quantopian.com.

thanks,
Jean