ETF Trend Prediction in Emerging Markets — Paper Reproduction

Reproduced Sagaceta-Mejía et al. (2024) using LASSO feature selection and MLP neural networks to predict ETF trends in Chile, Brazil, and U.S. equity markets. WQU MScFE coursework.

Paper: Sagaceta-Mejía et al., “An Intelligent Approach for Predicting Stock Market Movements in Emerging Markets Using Optimized Technical Indicators and Neural Networks,” Economics, vol. 18, no. 1, De Gruyter, 2024. DOI: 10.1515/econ-2022-0073  ·  PDF

Code: Google Colab  ·  Report  ·  Type: Paper Reproduction  ·  Context: WQU MScFE 600

Overview

Reproduced and extended a 2024 Economics paper applying LASSO-regularized feature selection and MLP neural networks to predict daily trend direction (up/down) in equity ETFs. Targets: ECH (Chile), EWZ (Brazil) as emerging market proxies, and IVV (S&P 500) as a developed-market benchmark.

All Python implementation was done independently.

Technical Approach

Feature Engineering

  • Constructed 27 technical indicators across four categories: trend (SMA, EMA, MACD, ADX), momentum (RSI, Stochastic, Williams %R, ROC), volatility (Bollinger Bands, ATR, standard deviation), and volume (OBV, MFI)
  • Expanded to 216 features via lag windows; binarized target as next-day trend direction

Feature Selection — LASSO Variants

  • Compared 8 LASSO-family methods: standard LASSO, Ridge, Elastic Net, Adaptive LASSO, and group variants
  • Evaluated feature stability across bootstrap samples using Jaccard distance (measures consistency of selected feature sets across resamples)
  • Selected(5) (LASSO selecting 5 features) identified as optimal: lowest Jaccard distance (highest stability) with ~9–10 features retained

Model

  • MLP classifier with k-fold cross-validation (k=10)
  • Selected(5) achieved ~2% accuracy improvement over the full 216-feature baseline
  • Reported per-ETF accuracy, precision, recall, F1; consistent gains in emerging market ETFs over developed market baseline

Key Findings

  • Emerging market ETFs (ECH, EWZ) showed stronger predictive signal from technical indicators than IVV — consistent with original paper’s hypothesis about inefficiency in emerging markets
  • Sparse feature selection (LASSO) generalizes better than full feature sets for daily trend classification
  • Jaccard stability criterion provides a principled method for selecting among regularization variants

Stack

Python · scikit-learn · pandas · NumPy · Matplotlib · LASSO / Elastic Net · MLP Classifier · k-fold CV

Connection to research: This project develops fluency in the methodological toolkit common to empirical ML research — feature selection under high dimensionality, regularization as a model selection criterion, stability analysis, and benchmarking across data regimes. The emerging-market framing also connects to applied econometrics and alternative data themes in my MSc coursework.