📈 economics

Does Complexity Improve Forecasting Accuracy? Evidence from Single and Multi-Layer LSTM Models for the BIST 100 Index

This study demonstrates that for forecasting the BIST 100 index using technical and macroeconomic variables, simpler single-layer LSTM architectures outperform more complex multi-layer models, suggesting that parsimony enhances generalization in financial forecasting.

Original authors: Fatih Mangir, Metehan Ortakarpuz, Abdul-Razak Bawa YUSSIF

Published 2026-10-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fatih Mangir, Metehan Ortakarpuz, Abdul-Razak Bawa YUSSIF

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of finance, predicting the future movement of a stock market index is a task that has long captivated economists and investors alike. It is a challenge rooted in the belief that while markets are chaotic, they are not entirely random; patterns hidden within past prices, economic reports, and global sentiment might offer a glimpse of what comes next. For decades, researchers have relied on statistical models to find these patterns, but the rise of powerful computers has introduced a new tool: artificial intelligence. Specifically, a type of computer program known as a neural network, designed to mimic the way the human brain processes sequences of information, has become a popular choice for this work. The prevailing assumption in this field has often been that more complex versions of these programs, with deeper layers of processing, should naturally lead to better predictions. After all, if a simple model can learn a little, a more intricate one should learn much more.

However, a recent study focused on the Borsa Istanbul 100 index, the primary benchmark for the Turkish stock market, challenges this assumption. The researchers, Fatih Mangir, Metehan Ortakarpuz, and Abdul-Razak Bawa Yussif, set out to test whether adding more layers of complexity to these prediction models actually improves their accuracy. They gathered over two and a half thousand days of trading data, stretching from early 2016 through the end of 2025, a period that included significant global and local economic shifts. To feed their models, they did not rely on price history alone. Instead, they combined daily closing prices with a rich mix of other information, including technical indicators that measure market momentum, global fear levels, the strength of the US dollar, and local economic factors like interest rates and unemployment. Before training their models, they used a mathematical technique to smooth out the noisy, erratic fluctuations in the data, ensuring the computer was learning from the underlying signal rather than random static.

The team then built two distinct families of models to see which would perform better. The first family consisted of simple, single-layer networks, while the second family featured more complex, multi-layered structures with multiple levels of processing. They trained these models to predict the next day's closing price, running the experiment thirty times for each configuration to ensure the results were not just a lucky fluke. The findings were clear and somewhat counterintuitive: the simpler models consistently outperformed the complex ones. The single-layer model with just ten processing units achieved the lowest average error, while the most complex models, including those with three layers, struggled to keep up. In fact, the most elaborate architecture produced the worst results of all, missing the mark by a wide margin.

This outcome suggests that in the specific context of forecasting this emerging market index, adding depth to the neural network did not help it learn better; instead, it may have caused the model to overcomplicate the task. The data, even after being cleaned and enriched with global and local variables, contained a structure that was best captured by a straightforward approach. The complex models appeared to be trying to learn patterns that simply did not exist, effectively memorizing the noise of the market rather than its true rhythm. The study indicates that for this type of financial forecasting, a parsimonious approach—one that uses the minimum necessary complexity to understand the data—is superior. The most accurate predictions came not from the most powerful machine, but from the one that was just complex enough to match the information available.

The researchers also noted an interesting trade-off in their results. While the simplest model was the most accurate on average, a slightly larger single-layer model with fifty units offered a different kind of benefit: stability. This version produced results that were nearly as accurate but far more consistent across different training runs, suggesting that for practical investment decisions, a model that is both reliable and accurate might be the ideal choice. Ultimately, the study serves as a reminder that in the realm of financial prediction, more is not always better. The best tool is not necessarily the most sophisticated one, but the one that is perfectly aligned with the nature of the data it is trying to understand.

Technical Summary: Does Complexity Improve Forecasting Accuracy? Evidence from Single and Multi-Layer LSTM Models for the BIST 100 Index

Problem Statement
The study addresses a critical methodological question in financial machine learning: whether increasing architectural complexity—specifically the depth of Long Short-Term Memory (LSTM) networks—necessarily enhances forecasting accuracy. While deep learning is often assumed to outperform simpler models due to greater representational capacity, the authors posit that in noisy financial environments, particularly within emerging markets like Turkey, excessive complexity may lead to overfitting. The research focuses on the BIST 100 index, a benchmark reflecting an emerging market exposed to domestic macro-financial instability and global shocks, to determine if parsimonious models generalize better than deeper architectures when the information structure of the data is limited.

Methodology
The study employs a rigorous, reproducible forecasting framework using 2,507 daily observations of the BIST 100 index from January 4, 2016, to December 31, 2025.

Data and Features: The target variable is the daily closing price. The input feature set includes technical indicators (MACD, RSI, ATR) and a multivariate set of global and macroeconomic variables (VIX, USDX, Turkish interest rates, unemployment, and consumer sentiment).
Preprocessing: To ensure temporal integrity and minimize information leakage, the data was split chronologically (80% training, 20% testing) before any processing. Missing values were imputed using forward/backward fill. Crucially, wavelet denoising (Haar wavelet with VisuShrink thresholding) was applied to the target variable to filter high-frequency noise, followed by Min-Max normalization derived strictly from the training set.
Sequence Construction: A five-day lookback window was used to create supervised learning sequences, capturing short-term market dependencies.
Model Architectures: Two families of LSTM models were compared:
- Single-Layer (SL): Configurations with 10, 30, 50, 100, 150, and 200 neurons.
- Multi-Layer (ML): Stacked configurations including two-layer structures (e.g., 100/50, 150/100) and a three-layer structure (100/50/20).
Training and Evaluation: Hyperparameters (optimizer, learning rate, batch size) were tuned via grid search for each architecture independently. To account for the stochastic nature of neural network training, each configuration was evaluated over 30 independent replications with random weight initializations. Performance was measured using Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and Pearson correlation. Statistical significance was assessed using Welch's two-sample t-test.

Key Results
The empirical findings demonstrate a non-monotonic relationship between model complexity and forecasting accuracy:

Superiority of Simplicity: The single-layer LSTM with 10 neurons achieved the lowest mean RMSE (231.32) and the highest Pearson correlation (0.9797), significantly outperforming all multi-layer architectures.
Degradation with Depth: Increasing the number of neurons in single-layer models beyond 50 led to performance deterioration, with the 200-neuron model showing the worst results. Similarly, multi-layer architectures generally exhibited higher error rates and lower correlation than their single-layer counterparts. The three-layer model (100/50/20) performed the worst among all configurations.
Stability vs. Accuracy Trade-off: While the 10-neuron model offered the highest accuracy, it showed higher variance across replications. The 50-neuron single-layer model emerged as a robust compromise, offering near-optimal accuracy with the greatest stability (lowest standard deviation in RMSE).
Statistical Significance: Welch's t-test confirmed that the performance difference between the best single-layer model (10N) and the representative multi-layer benchmark (150/100) was statistically significant ( $p < 0.001$ ), with the single-layer model providing a systematic accuracy advantage.

Key Contributions
The paper makes three primary contributions to the literature:

Multivariate Framework: It establishes a comprehensive forecasting framework for the BIST 100 that integrates technical indicators with global risk and domestic macroeconomic variables, processed through wavelet denoising.
Architectural Specificity: Unlike studies comparing different model families (e.g., ARIMA vs. LSTM), this research isolates the effect of architectural depth within the LSTM family, holding the information set and forecasting horizon constant.
Evidence for Parsimony: It provides empirical evidence from an emerging market context that deeper LSTM architectures do not inherently improve predictive performance. The study challenges the assumption that "deeper is better," suggesting that model capacity should be aligned with the informational structure of the data rather than expanded mechanically.

Significance and Claims
The authors conclude that in the context of volatile emerging market forecasting, greater architectural complexity does not necessarily improve forecasting accuracy. Instead, the results support a parsimonious approach where simpler, single-layer models may enhance generalization when the complexity of the model exceeds the information structure available in the data.

The study emphasizes that while multi-layer models may offer slightly more stability across training runs, they fail to capture the underlying market dynamics more effectively than compact architectures. The findings suggest that for financial time series with short lookback windows and specific noise characteristics, the "optimal" model is not the most complex one, but the one whose capacity is best matched to the data's informational content. The authors caution that these findings are specific to the experimental design (fixed lookback, specific predictors, and preprocessing) and do not universally preclude the utility of deeper models in scenarios with longer temporal dependencies or richer information sets.

Technical Summary: Does Complexity Improve Forecasting Accuracy? Evidence from Single and Multi-Layer LSTM Models for the BIST 100 Index

More like this