Paper: arXiv 2310.09903

Authors: Fatemeh Moodi, Amir Jahangard-Rafsanjani

Abstract

Stock price prediction is influenced by a variety of factors, including technical indicators, which makes Feature selection crucial for identifying the most relevant predictors. This study examines the impact of feature selection on stock price prediction accuracy using technical indicators. A total of 123 technical indicators and 10 regression models were evaluated using 13 years of Apple Inc. data. The primary goal is to identify the best combination of indicators and models for improved forecasting. The results show that a 3-day time window provides the highest prediction accuracy. Model performance was assessed using five error-based metrics. Among the models, Linear Regression and Ridge Regression achieved the best overall performance, each with a Mean Squared Error (MSE) of 0.00025. Applying feature selection significantly improved model accuracy. For example, the Multi-layered Perceptron Regression using Forward Selection improved by 56.47% over its baseline version. Support Vector Regression improved by 67.42%, and Linear Regression showed a 76.7% improvement when combined with Forward Selection. Ridge Regression also demonstrated a 72.82% enhancement. Additionally, Decision Tree, K-Nearest Neighbor, and Random Forest models showed varying levels of improvement when used with Backward Selection. The most effective technical indicators for stock price prediction were found to be Squeeze_pro, Percentage Price Oscillator, Thermo, Decay, Archer On-Balance Volume, Bollinger Bands, Squeeze, and Ichimoku. Overall, the study highlights that combining selected technical indicators with appropriate regression models can significantly enhance the accuracy and efficiency of stock price predictions.

Complexity vs Empirical Score

  • Math Complexity: 3.0/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper employs standard regression and machine learning models (e.g., Ridge, SVM, Neural Networks) and wrapper feature selection methods without novel mathematical derivations or advanced theoretical proofs. However, it demonstrates strong empirical rigor through a rigorous evaluation on 13 years of real Apple Inc. data, using 123 technical indicators, 10 models, multiple time windows, and comprehensive error metrics (MSE, MAE, etc.), with reported percentage improvements in accuracy.

Research Flowchart

  flowchart TD
  A["Research Goal: Optimize Stock Price Prediction<br>using Feature Selection & Regression"] --> B["Data Input<br>13 Years of Apple Inc. Data"]

  B --> C{"Methodology Steps"}
  C --> D["Feature Generation<br>123 Technical Indicators"]
  C --> E["Feature Selection<br>Forward & Backward Selection"]
  D --> E
  E --> F["Model Evaluation<br>10 Regression Models"]

  F --> G{"Key Findings & Outcomes"}
  G --> H["Optimal Time Window: 3 Days"]
  G --> I["Best Models (MSE 0.00025):<br>Linear & Ridge Regression"]
  G --> J["Impact of Feature Selection:<br>56-76% Accuracy Improvement"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes using regression models combined with forward and backward feature selection on technical indicators to improve Apple stock price prediction.

Source sentence
In this study, we use machine learning techniques, specifically regression models combined with feature selection methods, to improve prediction accuracy.

Assumes 4

  • The data is split into training and testing sets using a 70‑30 ratio. data
    Source sentence
    For both parts of the dataset, a 70-30 ratio is applied to split the data into training and testing sets.
  • K‑fold cross‑validation provides a reliable performance estimate despite limited sample size. statistical
    Source sentence
    To mitigate this issue, K-fold cross-validation is employed.
  • A 3‑day time window is appropriate for converting the series into a supervised learning problem. model
    Source sentence
    Here, like Ji et al. [48], a 3-day time window is intended for the data set and a new data set is created.
  • Technical indicators computed from historical price data are sufficient predictors. market
    Source sentence
    Various technical indicators (listed in Table 1) are computed, and a new dataset is created that includes these technical indicators alongside the stock data.

Data

  • Apple Inc. (AAPL) historical price data · Yahoo Finance market-prices public
    01/01/2010 to 31/12/2022 · Apple Inc. stock
    Source sentence
    Figure 2 illustrates the process. First, the stock data of Apple Inc. (AAPL) is collected from Yahoo Finance.

Recognised sources: Yahoo Finance

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Linear Regression and Ridge Regression achieved the lowest MSE of 0.00025, the best overall performance. performance
    Source sentence
    Among the models, Linear Regression and Ridge Regression achieved the best overall performance, each with a Mean Squared Error (MSE) of 0.00025.
  • A 3‑day time window provided the highest prediction accuracy among tested windows. predictive-accuracy
    Source sentence
    The results show that a 3-day time window provides the highest prediction accuracy.
  • Forward Selection improved MLP regression MSE by 56.47%, SVR by 67.42%, Linear Regression by 76.7%, and Ridge Regression by 72.82%. performance
    Source sentence
    MLPSF outperformed MLP by 56.47%, SVRSF outperformed SVR by 67.42%, LRSF outperformed Linear Regression by 76.7%, RISF outperformed Ridge Regression by 72.82%.
  • The most effective technical indicators were Squeeze_pro, Percentage Price Oscillator, Thermo, Decay, Archer On‑Balance Volume, Bollinger Bands, Squeeze, and Ichimoku. statistical
    Source sentence
    Several technical indicators were consistently selected across different regression models, including: Squeeze_pro, Percentage Price Oscillator, Thermo, Decay, Archer On-Balance Volume, Bollinger Bands, Squeeze, and Ichimoku.
  • Sequential Backward Selection was slower and more memory‑intensive than Sequential Forward Selection. other
    Source sentence
    The Sequential Backward Selection (SBS) algorithm was observed to be slower than Sequential Forward Selection (SFS) due to its higher memory demands at the initial stage.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 21 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.