Logistic Regression vs KNN: Python Stock Trading Experiment

In this article the author, a Python enthusiast, reproduces a quantitative trading strategy from a Tsinghua University book, replacing the original K‑Nearest Neighbors model with a logistic regression model to compare performance.

import pandas as pd
import pandas_datareader.data as web
import numpy as np
from datetime import datetime, timedelta

Three years of daily data for Kweichow Moutai (ticker 600519.SS) are retrieved from Yahoo Finance.

end = datetime.date.today()
start = end - timedelta(days=365*3)
owB = web.DataReader('600519.ss', 'yahoo', start, end)

Feature engineering creates two price‑difference columns and a binary target indicating whether the next day’s closing price is higher (1) or lower (‑1).

owB['open-close'] = owB['Open'] - owB['Close']
owB['high-low'] = owB['High'] - owB['Low']
owB['target'] = np.where(owB['Close'].shift(-1) > owB['Close'], 1, -1)
owB = owB.dropna()

The dataset is split into features x and label y, then into training and test sets (80 % training).

x = owB[['open-close', 'high-low']]
y = owB['target']
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
x_train, x_test, y_train, y_test = train_test_split(x, y, train_size=0.8)

A logistic regression model is trained and evaluated.

lr = LogisticRegression()
lr.fit(x_train, y_train)
print(lr.score(x_train, y_train))
print(lr.score(x_test, y_test))

The training accuracy is 0.5439, while the test accuracy drops to 0.5137, slightly lower than the KNN benchmark reported in the book.

The author notes that the modest performance may stem from noisy data or the limited predictive power of simple price‑difference features for stock movements.