Machine Learning using Neural Networks | Educational

13 oct. 2023

The script provided is a comprehensive illustration of how to implement and execute a simplistic Neural Network (NN) on TradingView using PineScript.

It encompasses the entire workflow from data input, weight initialization, implicit neuron calculation, feedforward computation, backpropagation for weight adjustments, generating predictions, to visualizing the Mean Squared Error (MSE) Loss Curve for monitoring the training phase.

In the visual example above, you can see that the prediction is not aligned with the actual value. This is intentional for demonstrative purposes, and by incrementing the Epochs or Learning Rate, you will see these two values converge as the accuracy increases.

Hyperparameters:
Learning Rate, Epochs, and the choice between Simple Backpropagation and a verbose version are declared as script inputs, allowing users to tailor the training process.

Initialization:
Random initialization of weight matrices (w1, w2) is performed to ensure asymmetry, promoting effective gradient updates. A seed is added for reproducibility.

Utility Functions:
Functions for matrix randomization, sigmoid activation, MSE loss calculation, data normalization, and standardization are defined to streamline the computation process.

Neural Network Computation:
The feedforward function computes the hidden and output layer values given the input.
Two variants of the backpropagation function are provided for weight adjustment, with one offering a more verbose step-by-step computation of gradients.

A wrapper train_nn function iterates through epochs, performing feedforward, loss computation, and backpropagation in each epoch while logging and collecting loss values.

Training Invocation:
The input data is prepared by normalizing it to a value between 0 and 1 using the maximum standardized value, and the training process is invoked only on the last confirmed bar to preserve computational resources.

Output Forecasting and Visualization:
Post training, the NN's output (predicted price) is computed, standardized and visualized alongside the actual price on the chart.

The MSE loss between the predicted and actual prices is visualized, providing insight into the prediction accuracy.

Optionally, the MSE Loss Curve is plotted on the chart, illustrating the loss trajectory through epochs, assisting in understanding the training performance.

Customizable Visualization:
Various inputs control visualization aspects like Chart Scaling, Chart Horizontal Offset, and Chart Vertical Offset, allowing users to adapt the visualization to their preference.

-------------------------------------------------------

The following is this Neural Network structure, consisting of one hidden layer, with two hidden neurons.



Through understanding the steps outlined in my code, one should be able to scale the NN in any way they like, such as changing the input / output data and layers to fit their strategy ideas.

Additionally, one could forgo the backpropagation function, and load their own trained weights into the w1 and w2 matrices, to have this code run purely for inference.

-------------------------------------------------------

While this demonstration does create a “prediction”, it is on historical data. The purpose here is educational, rather than providing a ready tool for non-programmer consumers.

Normally in Machine Learning projects, the training process would be split into two segments, the Training and the Validation parts. For the purpose of conveying the core concept in a concise and non-repetitive way, I have foregone the Validation part. However, it is merely the application of your trained network on new data (feedforward), and monitoring the loss curve.

Essentially, checking the accuracy on “unseen” data, while training it on “seen” data.

-------------------------------------------------------

I hope that this code will help developers create interesting machine learning applications within the Tradingview ecosystem.

15 oct. 2023

Notes de version

+ License