Neural Networks in Trading - DNN model - Return calculation

AI & Machine learning
1

In the Strategy return section, it is using cumulative sum instead of product:



"…Once we calculate the percentage returns of the model, we will take a cumulative sum of all these returns on the test data to measure the overall performance of the model…"

final_return = np.nancumsum(strategy_returns)*100



Shouldn't it be (returns+1).cumprod() where the return is already multiplied with trend or signal.

2

Hi Soumaneel Konar,



Yes, you are correct cumprod gives more accurate results than np.nancumsum. Thanks for pointing out the error. We will make changes to our notebook also.



Thanks,

Danish

3

Thank you.



One more thing plz look into in the same notebook: DNN Trading Strategy Code



Labels / Signals are created using shift(-5) but the strategy return has been calculated using shift(-1).

Is this correct? Because prediction will correspond to the labels that have been created right?

If the labels are based on data 5 days ahead, the prediction will also be 5 days ahead. Isn't it?



Below is the code from the notebook:



Label / Signal creation:

y = np.where(data.Close.shift(-5) > data.Close, 1, 0)



Return calculation:

market_returns = pd.Series(data.Close.shift(-1).iloc[-test_size:].pct_change().values)

strategy_returns = pd.Series(predict_trend*market_returns)

4

We are using a Neural network for predicting signals and we are using shift(-5) for generating target values. Using shift(-5) can help us recognize patterns further ahead. But in the case of calculating strategy returns, we can only use shift(-1) because we want to calculate returns in each consecutive period.