Backtesting multiple trading strategies against a single asset

[Sanjan611]

I'm using Portfolio.from_signals to backtest multiple trading strategies against a single asset and then compare their Sharpe ratios. Currently, my Portfolio's entries parameter is a <DataFrame> E where each column of E represents the entry signals of one of the strategies. And then I have group_by = E.keys() so that portfolio stats can be calculated grouped by that strategy. Similarly for the exit signals.

Would this be the suggested approach using vectorbt for such a task? This works as expected, but I'm not sure if using a massive DataFrame for entry and exit signals (assuming I'm backtesting a very very large number of strategies) utilises the performance benefits that this library offers. Any thoughts on this would be appreciated!

[polakowo]

@Sanjan611 yes, this is the way. Boolean arrays don't take much memory anyway. If any of your information pieces (such as price, size, etc.) is the same for all strategies, you can pass it as a pandas series and it will broadcast to a frame in a lazy way (without materializing, just by smart indexing - effectively no further memory occupation). So, for example, doing vbt.Portfolio.from_signals([1, 2, 3], True, False, size=10) is the same as doing vbt.Portfolio.from_signals([1, 2, 3], [True, True, True], [False, False, False], size=[10, 10, 10]), but uses much less memory. If you pass a pandas series, it's assumed to be an element per row. If you pass a 1-dim NumPy array, it's assumed to be an element per row/column, depending upon whether all inputs are 1-dim. If you pass a const, it's assumed to be per element. Broadcasting is done using vectorbt's own broadcasting mechanism (here). If memory is the main bottleneck, you can port your strategy to Portfolio.from_order_func to use sparse matrices or alike, but you would need to implement all signal logic by yourself.

[Sanjan611]

I see, thanks!