Training a simple weekly resistance model for SPY (Part 1)

This is my first attempt at developing a machine learning model for quantitative analysis and trading. I don't know what to expect, but I do know I will encounter a lot of challenges along the way. This model will be very simple and likely under/overfit, inaccurate, and not useful for trading.

Objective: Create a simple XGBoost model to predict weekly resistance levels for SPY.

Data

• SPY daily price data (20 years)

Preprocessing the data

1. Columns: datetime, high, low
2. Round the high and low to ticks (1% of the price to 1 significant digit)
3. Add a column for the day of the week (0-4)

Preparing the dataset

• Features: array of high and low prices
• 10 days of data (2 weeks from Monday to Friday)
• Exclude data if it doesn't have 10 days of data
• Target: next week's high and low price for the week

Model

Simple XGBRegressor model with 100 estimators, 0.1 learning rate and default parameters.

Data is also scaled using StandardScaler.

I know this is very bad, but I want to start simple and then improve it.

A model is trained for the high and low prices separately.

Results

XGBoost Regressor Performance for High Data:
Train MSE: 0.3408
Test MSE: 26.6598
Train R2: 1.0000
Test R2: 0.9980

XGBoost Regressor Performance for Low Data:
Train MSE: 0.2902
Test MSE: 56.1965
Train R2: 1.0000
Test R2: 0.9958

Analysis

The model looks great, right? No.

$R^2$ for training is perfect ($1.0$) and test is very high ($0.9980$ and $0.9958$) is very high too.

But, I know this is very bad. The model is likely overfitting and not generalizing well to be useful for trading.

Visualizing Error

NEXT: Training a simple weekly resistance model for SPY (Part 2)