Walk Forward Optimization in Python with Backtesting.py

⭐ Code:

⏰ Timestamps:

00:00 - Introduction
03:47 - Data Sourcing / importing
05:50 - Writing our Momentum strategy
20:10 - Walk forward Function
36:00 - Plotting our out of sample backtests
44:55 - Plotting combined equity curve
56:57 - Plotting sample and validation data ranges

[pethiraj]

Amazing content and with this high quality content, wish you the very best and soon million subscribers!

[VaidaBogdan]

Loved the video, exactly what I needed. Would love to get access to your notebooks too.

[christolabuschagne]

I've been binge watching your stuff for a few days now. This is a great video with a solid idea. My only problem is that my strategy allows for multiple trades to be opened, but as soon as the week ends, it closes all my open trades, and this causes my performance to tank a bit. If I run it over full data set I get very good results, but slicing it like this really kills my return %. Not sure if something can be done about it though...

[samyoe]

I use an ML model that splits the data into training and validation sets within the strategy itself. So I just got rid of the validation backtest within the walk_forward function, and it worked fine.

[TIMHOK_]

Thank you for your video!
After testing my brilliant strategy with walk-forwards i came to conclusion that its not that brilliant anymore huh.
I now feel like backtesting any strategies over big amounts of klines (ie long periods of market data) is delusional and walk-forward is the only way to prove it working.
If strategy performs poorly, do you think optimization over params such as lookback_bars and validation_bars worth doing or its better to move on and get a better stategy?

Thanks for your work. Is any way get the code you show in the video?

[richardchu]

Any chance we can add a Monte Carlo and MAE/MFE analysis please?

[VaidaBogdan]

I'm curious of your view regarding what would be a very good backtesting methodology. E.g. kfold validation, walk-forward optimization (anchored or not)... there are multiple ways to look at it and you've mentioned on your website that you're into the hard stuff... I'm curios of your view on what a default / go-to method would do.

For context: I started out splitting 60-20-20 and quickly realized I couldn't do that because I would, for example, train on a sideways market, validate on an upside and test on an upside. Then used kfold and split the timerange into 5 folds, each with its own training and testing set. That worked well enough but on higher timeframes I couldn't get enough trades to validate the results. This is my default framework but I sometimes use a walk-forward that is anchored... but I wanted to know how more experienced quants do it. If the 5 fold approach is better or worse than an unanchored walk forward or than an anchored one or if there is another way to look at this that is escaping me.

[dgarciagud]

Are going to going to make it available somewhere?

[amedmostafaabuhatab]

I have a question, in the walk forward function, why bt.optimize not bt_training.optimize!

[SaraStemPro]

Thank you so much for the information, it's very useful. Could you use the split data from sklearn?

[samueltamayogaviria]

Hello Chad. Thanks for your videos, they have served very well for my formation as an algorithmic trader. I have a question, if I am not calculating indicators for other timeframes, do I still need the "warmup_bars"?

[hwtang]

Nice content, Chad. I am new for back testing, I have no idea how to rewrite the walk forward if using different time frame with 2 moving average crossover strategy with optimisation. Thanks.

[FabioArcari-nl]

Hi, is there a way to download the .py you're explaining? Thanks

[xramonz]

I don't understand how to choose parameters after such optimizations. For example we have bunch of results which is obtained with different strategy parameters. What would we choose for further trading in real market? Which parameters are best?

[aarondelarosa]

Hey Chad. I can't install backtesting on terminal. What do I do? I'm using Jupyter Notebook VSC (Visual Studio Code).

[OlufemiJeromeAjakaye]

Your view platform not yet clear for the viewers to see d zoom up.

[re_di_roma_is_back]

Hi. My Jupiter says "No module named backtesting"

[kanaya]

How can used metatrader data here, l?😁

[AccOunt-cbtc]

I had a problem with the optimized parameters for the backtests on the validation sets of data all being the same when printing them from the _strategy in the stats results.
When you tested it in your video you only checked a single _strategy result to see if it was different from the defaults, but not if they were different from each other so I don't know if you had the same problem, I think I copied your code exactly in a new jupiter notebook and still had the same problem, anyone else have this problem?

edit:

I fixed it like this but I don't know what the original problem was.

def walk_forward(strategy, data_full, warmup_bars, lookback_bars, validation_bars, cash=10_000_000, commission=0.002):
stats_master = []
params_master = []

for i in range(lookback_bars + warmup_bars, len(data_full) - validation_bars, validation_bars):
training_data = data_full.iloc[i - lookback_bars - warmup_bars:i]
validation_data = data_full.iloc[i - warmup_bars:i + validation_bars]

bt_training = Backtest(training_data, strategy, cash=cash, commission=commission)
stats_training = bt_training.optimize(
small_threshold=list(np.arange(0, 1, 0.1)),
large_threshold=list(np.arange(1, 3, 0.2)),
maximize="Equity Final [$]"
)

optimized_strategy = stats_training._strategy
strategy.small_threshold =
strategy.large_threshold =

bt_validation = Backtest(validation_data, strategy, cash=cash, commission=commission)
stats_validation = bt_validation.run()


params_master.append({
"small_threshold": optimized_strategy.small_threshold,
"large_threshold":
})

return stats_master, params_master

lookback_bars = 28*400
validation_bars = 7*400
warmup_bars = 16*24

stats, params = walk_forward(
MomentumStrategy,
data,
lookback_bars=lookback_bars,
validation_bars=validation_bars,
warmup_bars=warmup_bars
)
stats

params[0], params[1], params[2], params[3], params[4], params[5], params[6], params[7], params[8], params[9], params[10]