ETF Forecast: Application of Innovative Machine Learning Models

in the Field of ETF Forecasting

Tong Zhang

Department of Business, University of New South Wales, Kensington, Sydney, Australia

Keywords: ETF Prediction, Machine Learning Models, Innovative Algorithm, Financial Technology.

Abstract: With global exchange-traded fund (ETF) assets exceeding $15 trillion, ETFs are becoming increasingly

significant in the global financial system. Specifically, because of the unpredictability of the outside world

in recent years, investors have been more interested in low-cost, low-risk, and highly transparent investments,

which is reflected in the increasing scale of passive investment ETFs and the increase in research in the field

of ETF prediction. This study examines how machine learning models are now being used in the field of ETF

prediction and chooses innovative machine learning models from the previous two years to review. Including

the superposition of common models, the combination of traditional financial models and deep learning

models, etc., the results show that these innovative combinations can significantly improve the effectiveness

of ETF predictions. Researchers who wish to develop innovative machine learning models for ETF prediction

will benefit from this study's understanding of current findings and possible research directions.

1 INTRODUCTION

On January 23, 1993, Standard & Poor’s Depositary

Receipt (SPDR) launched the first ETF in the United

States, named after its underlying index. SPDR S&P

500 is the largest open-end index fund in the world

(Liebi, 2020). Like mutual funds and index portfolios,

exchange-traded funds (ETFs) were created by

assembling a group of stocks. Existing ETFs are a

further extension of index funds, with advantages

including ease of trading, tax benefits, and significant

cost-effectiveness (Joshi & Dash, 2024). From the

late 1990s to the early 21st century, ETFs began to

expand globally. Between 2003 and 2023, the total

assets under ETF control increased from 204.3 billion

to 11,507 billion USD (Joshi & Dash, 2024; Statista

Research Department, 2024).

Since the first ETF appeared on the market in

1993, the ETF market has developed rapidly. In 2003,

there were only 276 ETFs in the world. As of 2023,

this number has grown to 10,319 (Statista Research

Department, 2023). The rapid development of ETFs

in recent years demonstrates the further

intensification of global investors' interest in ETF

investment, a trend that began after the 2008 financial

crisis (EPFR, 2021). ETFs are popular among

https://orcid.org/0009-0008-2667-0287

investors for their high liquidity, low fees and lower

volatility compared to stocks (Liebi, 2020; Joshi &

Dash, 2024). A 2022 survey of 60 executives

worldwide by PricewaterhouseCoopers (PwC)

pointed out that the strong resilience and growth

potential of ETFs in resisting risks have been further

highlighted during COVID-19

(PricewaterhouseCoopers, 2021). With the

contribution of unprecedented capital inflows, a

substantial number of new entrants and diverse

product innovations and distribution opportunities,

the activity, diversity and innovation of the ETF

market are constantly increasing, and it has become

an important part of the global asset and wealth

management field.

Today, nearly 50% of investment in the United

States comes from ETFs, and the size of the ETF

market in the United States accounts for about 70%

of the overall ETF market size, leading the world

(Joshi & Dash, 2024). In particular, in August 2019,

passive investment exceeded active investment in the

US ETF market for the first time (Joshi & Dash,

2024).

This further shows that investors are increasingly

pursuing stable returns, which has promoted the

Zhang, T.

ETF Forecast: Application of Innovative Machine Learning Models in the Field of ETF Forecasting.

DOI: 10.5220/0013700100004670

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Conference on Data Science and Engineering (ICDSE 2025), pages 497-501

ISBN: 978-989-758-765-8

497

further in-depth application of machine learning

algorithms ETF performance forecasting.

Stock price forecasting is the central challenge at

the nexus of computer science and finance (Yüksel,

2023). The well-established efficient market theory

suggests that financial markets efficiently process

information, ensuring that key data is swiftly and

accurately incorporated into stock valuations.

Consequently, investors struggle to achieve

consistent returns above the market average without

engaging in manipulative practices. While this

perspective may seem pessimistic, it underscores the

inherent difficulty of predicting market movements.

Furthermore, the financial sector plays a vital role in

economic growth, attracting extensive research and

innovation from scholars worldwide.

2 EXISTING LITERATURE

ETF, as an investment tool that has flourished in the

past few years, has attracted the attention of an

extensive amount of researchers and investors. At

present, most of the research in this field is to verify

the model validity of a single machine learning model

for a single or a few ETF markets.

2.1 Previous Literature

To validate the LSTM model's reliability in

forecasting the return direction and particular ETF

prices, Horst employed ETFs from five distinct

marketplaces (Horst, 2022).

The study found that the Root Mean Squared

Error (RMSE) values of different industries vary

greatly, so the accuracy of the model is strongly

correlated with the choice of industry, with the real

estate industry performing the best and the energy

industry performing the worst. Adding or reducing

the number of ETFs will also greatly affect the

prediction results.

Gowani and Kanjiani also chose the LSTM model

and tested it in 9 different industries and more than

2,200 Vanguard industry ETFs (Gowani & Kanjiani,

2024). The difference is that the researchers chose R-

squared value to evaluate the effectiveness of the

model. The results became more optimistic, with R-

squared values exceeding 0.68 in all industries and

reaching 0.9095 in the energy industry, showing its

strong predictive ability.

Astuy et al. selected Extreme Gradient Boosting

(XGB), Neural Prophet model, Extra Trees (ET), and

K-nearest Neighbors (KNN) for three ETFs

(Sagarzazu Astuy, 2022). The results show that the

prediction effects of different models vary greatly.

When making investment decisions, it is advisable to

employ multiple models and incorporate diverse

predictive outcomes to enhance accuracy and

robustness.

Between January 1, 2006, and October 31, 2023,

Wang & Zhang chose ten distinct ETFs from the S&P

500 (Wang & Zhang, 2024). The model selection

includes Decision Tree (DT), Gaussian Naive Bayes

(GNB) and Neural Networks (NN).

The concentration of this research is on

forecasting the performance of several ETFs in the

future. Based on the forecasts, trading strategies are

then developed, and the actual returns are monitored.

Below market performance, that is, the short-selling

trading strategy can obtain returns that exceed the

return of holding the S&P 500 alone.

Through the investigation of existing research, it

is found that machine learning models are widely

used in this field, whether predicting the direction of

returns, specific prices or market performance, but

most of them use original models, and different

models have different performances in different

markets. Therefore, this study pays special attention

to the innovation and universality testing of the

model, hoping to improve the effectiveness of

prediction.

2.2 Innovative Algorithm Model

Chang et al. took the “Yuanta/P-shares Taiwan Top

50 ETF (Exchange Traded Fund)” (ETF50) as the

research object(Chang et al., 2024). ETF 50 is an

open-end exchange-traded index fund in the Taiwan

market, which aims to provide investors with

investment returns similar to the performance of the

stocks of the top 50 companies in Taiwan by market

value. The trading code is 0050. ETF50's holdings

cover multiple industries and fields such as

semiconductors, financial insurance, and electronic

components. Among them, TSMC, as the leading

company in global semiconductor manufacturing,

plays a pillar role in Taiwan's economy.

The typical stock prediction model analysis

method relies heavily on past data and assumes that

future market patterns will stay within a known

range(Chang et al., 2024). However, in recent years,

this initial assumption has been broken by unforeseen

disruptive events, including the public health crisis

around 2020, the economic dispute between China

and America in 2018, and the American financial

turmoil in 2011.

ICDSE 2025 - The International Conference on Data Science and Engineering

498

These events have disrupted and significantly

shattered production activities and affected the

regular operation of numerous industries, underlining

the gravity of the situation. Therefore, the researchers

adopted a new calibration process "Short-Term Bias

Compensation" (STBC), aiming to further calibrate

LSTM model predictions. This method could

minimize the volatility influence impact of sudden or

extreme events on forecasting accuracy.

Additionally, fuzzy rules evolved via genetic

algorithms (GA) are employed for optimizing trading

strategies.

The dataset covered ETF trading data from 2003

to 2020, and black swan events are deliberately

included to assess model precision. By calculating

daily forecast errors and comparing the daily

anticipated amounts with the actual data, the

prediction's short-term bias (STB) was first

estimated. The anticipated price for the following day

will be automatically adjusted if the STB rises above

a predetermined level. The STBC parameters are

further optimized by using genetic algorithms (GAs).

Simultaneously, trading strategies and stock buying

and selling signals were determined by genetic fuzzy

systems (GFS). According to the findings, STBC can

cut the prediction error by over 90%.

Shih et al. focused on exploring how to combine

traditional financial models (such as the Fama-French

three-factor model, the capital asset pricing model

(CAPM)) and artificial neural network architectures

applicable to handling time series, including LSTM,

ANN, GRU, CNN and their variants, to obtain better

results in ETF daily return prediction (Shih et al.,

2024). The multi-factor market model and the Taiwan

Economic Journal (TEJ) provided the data for this

Python-based analysis. Daily returns of six ETFs

listed in Taiwan: Yuanta Taiwan 50 (0050), Yuanta

Mid-Cap 100 (0051), Yuanta Electronics (0053),

Yuanta S&P Custom China 50 (0054), Yuanta MSCI

Taiwan Financial (0055), and Yuanta Taiwan

Dividend Plus (0056) were chosen as the dependent

variables. The time range spans from 2010 to 2020.

The study is divided into three parts. First, the

Fama-French three-factor model and the deep

learning algorithm's daily return prediction effects are

compared. The contrast of the linear traditional model

and the nonlinear artificial neural network is one of

the distinctive study features.

The findings demonstrated that the nonlinear

ANN combined with the CAPM, the Fama-French

three-factor model and the Fama-French five-factor

model yielded significantly stronger predictive power

than regression-based approaches. First and second

place went to the ANN combination with the Fama-

French three-factor and Fama-French five-factor

models, respectively. Other commonly used artificial

neural networks, LSTM and GRUs, were further

studied. The prediction effects of these two models

were compared with the CAPM, the Fama-French

three-factor model and the Fama-French five-factor

model. The study demonstrated that the Fama-French

three-factor model in association alongside any

artificial neural network performed better than the

Fama-French five-factor model and the CAPM in

conjunction with other models. Out of all the

combinations that contained the Fama-French three-

factor model, the combination of LSTM and it yielded

the lowest mean absolute error (MAE) value.

Furthermore, the Fama-French three-factor

model's MAE values were all lower than ANN's when

combined with LSTM and GRU. Moreover, the study

began to add other variables to explore better

prediction results. The added factors include

Momentum Factor, Investment Factor, Profitability

Factor, Dividend Yield Factor, Long-Term Reversal,

and Short-Term Reversal. The results showed that the

combination of the Fama-French three-factor model

performs well regardless of whether the Short-Term

Reversal factor is added. The researchers also added

CNN but found that the prediction error increased

after adding CNN except in 2016 and 2019.

Therefore, the model using hybrid or stacked

networks has not effectively improved the ability to

explain daily prices. The study pointed out that the

combination of the Fama-French three-factor model

and LSTM was the most effective way to predict daily

returns, providing excellent prediction accuracy.

Using the historical return data of its constituent

equities, this study (Piovezan, de Andrade Junior, &

Ávila, 2024) suggested a machine learning-based

prediction strategy for the direction of ETF returns.

The study used information from five reference

markets to compare the outcomes with buy-and-hold

and naive prediction strategies. Regression models

(linear regression, ridge regression, extreme gradient

boosting (XGBoost), lightweight gradient boosting

(LightGBM), and classification models (logistic

regression, support vector machine (SVM), naive

Bayes (GaussianNB), K-nearest neighbour (KNN),

and random forest) were among the machine learning

models that were employed. Each model was studied

and applied to five underlying indexes. BOVA11,

SPY, DAXEXx, MAXIS, and ISF ETFs reflect the

Ibovespa, S&P500, DAX, NIKKEI, and FTSE

indexes. The 12 largest constituent stocks that make

up each index were selected, including their daily

closing prices from January 1, 2012 to January 25,

2022, and the entire data set contains about 2,500

ETF Forecast: Application of Innovative Machine Learning Models in the Field of ETF Forecasting

499

trading days. The historical data of these constituent

stocks are used as the input data of the model to

predict the return direction of the ETF on the

following trading day.

Python 3.6.5 is the programming language, and

Anaconda is the integrated development environment

(IDE) used for programming. In order to ensure the

robustness of the algorithm, the NumPy function was

first used to calculate the logarithmic return and

preprocess the data.

The data set was divided in half, taking into

account the previous 1,000 trading days, and the

distribution ratio between testing and training was set

at 60/40%. The daily ETF return (t+1) was reversely

distributed back to the returns of its 12 constituent

stocks on the day before (t) to train the algorithm.

This enabled the algorithm to ascertain the

relationship between the ETF's current day returns

and the returns of its constituent equities the day

before. A binary classification (0, 1) represented the

ETF's return on the current day (t+1). If the predicted

return was greater than zero, it was classified as 1, and

if it was less than 0, it was classified as 0. Next, a

machine learning model was imported to implement

this method. Sharpe index, profit factor and

maximum drawdown were used as financial metrics,

and mean square error, root mean square error and

mean absolute error are used as error metrics.

The results showed that among the two control

strategies of buy and hold and naive forecast, SPY

achieved the highest return, which can be explained

by the higher linearity of growth of SPY constituents,

while BOVA and NIKKEI had lower returns, and

DAX and FTSE had downward returns. The machine

learning model performed better than buy and hold on

the Sharpe ratio and profit factor, except for BOVA

in the classification model and FTSE in the

classification model. The classification model and

LSTM model showed the lowest drawdown value in

general.

According to an analysis of the computational

efficiency of errors and scores, most regression

models had better scores and fewer prediction

mistakes than naive forecasts. Machine learning

models typically outperformed the naïve control

approach and occasionally outperformed the buy and

hold strategy when evaluating logarithmic returns.

The highest-performing models were KNN and SVM,

whereas the worst-performing models were LGBM

regression and LGBM classifier. In conclusion,

machine learning is an optional tool for financial data

prediction. By forecasting the direction of returns,

most models outperform buy and hold strategies

regarding returns, errors, and scores.

3 CONCLUSIONS

This study selected a relatively novel model

combination and research method in ETF prediction

using machine learning Algorithms. This paper paid

special attention to the superposition and combination

of models, as well as the universality of models, and

tried to find better ways to improve the effectiveness

of ETF prediction. The results showed that STBC can

reduce nearly 90% of prediction errors, and the

combination of Fama-French three-factor model and

LSTM can significantly improve the accuracy of

prediction. It was worth noting that the effectiveness

of different models in different ETF markets varies,

so when choosing, it was still necessary to pay

attention to their adaptability to specific ETFs.

Currently, there are few studies on innovative models

in the field of ETF forecasting, and their universality

needs to be further proven. Most studies only focus

on the application of one or several models to

individual ETFs. Nevertheless, in the field of stock

prediction, the number of similar innovative studies

is relatively rich, which may be related to the

difficulty of obtaining data sets. However, relatively

speaking, the ETF market is still in the development

stage, and ETF is still a young investment tool. It is

reasonable that there are few related studies.

Therefore, the researcher suggests that more public

data sets belonging to ETFs should be established in

the future to help other researchers reduce the

difficulty of obtaining data sets and save a certain

amount of data preparation time. Furthermore, the

algorithmic process of deep learning models is still in

a "black box" state. Researchers can visualize the

model training process in future research or choose to

use related Explainable AI (XAI) methods to improve

the transparency and interpretability of the algorithm.

This study focuses on innovation and

comprehensiveness beyond conventional methods,

and opens up new ideas for researchers in the field of

ETF prediction. This paper hope that researchers will

fill the gap in this research field in the future, continue

to explore innovative combinations and optimizations

of machine learning models, and improve prediction

accuracy.

REFERENCES

Chang, R. I., Wang, C. H., Wei, L. C., & Lu, Y. F. (2024).

LSTM with short-term bias compensation to determine

trading strategy under black swan events of Taiwan

ETF50 stock. Applied Sciences, 14(18), 8576.

ICDSE 2025 - The International Conference on Data Science and Engineering

500

EPFR. (2021, November 11). Data reports global ETF

assets exceed $10 trillion. EPFR. Retrieved from

https://epfr.com/insights/press-release/epfr-data-

reports-total-global-etf-assets-exceed-10-trillion/

Gowani, R., & Kanjiani, Z. (2024). Advanced LSTM neural

networks for predicting directional changes in sector-

specific ETFs using machine learning techniques. arXiv

preprint arXiv:2409.05778.

Horst, J. (2022). Machine learning for exchange-traded

fund price predictions (Bachelor's thesis, University of

Twente).

Joshi, G., & Dash, R. K. (2024). Exchange-traded funds and

the future of passive investments: A bibliometric

review and future research agenda. Future Business

Journal, 10(1), 17.

Liebi, L. J. (2020). The effect of ETFs on financial markets:

A literature review. Financial Markets and Portfolio

Management, 34(2), 165-178.

Li, H., Shen, Y., & Zhu, Y. (2018, November). Stock price

prediction using attention-based multi-input LSTM. In

Asian Conference on Machine Learning (pp. 454-469).

PMLR.

Piovezan, R. P. B., de Andrade Junior, P. P., & Ávila, S. L.

(2024). Machine learning method for return direction

forecast of exchange-traded funds (ETFs) using

classification and regression models. Computational

Economics, 63(5), 1827-1852.

PricewaterhouseCoopers. (2021). ETFs 2026: The next big

leap. PwC. Retrieved from

https://www.pwc.com/gx/en/industries/financial-

services/publications/etf-2026-the-next-big-leap.html

Sagarzazu Astuy, E. (2022). Estimación del retorno y riesgo

esperado de un Exchange Traded Fund (ETF) aplicando

modelos de machine learning (ML).

Shih, K.-H., Wang, Y.-H., Kao, I.-C., & Lai, F.-M. (2024).

Forecasting ETF performance: A comparative study of

deep learning models and the Fama-French three-factor

model. Mathematics, 12(19), 3158.

Statista Research Department. (2023). Number of

exchange-traded funds (ETFs) worldwide from 2003 to

2023. Statista. Retrieved from

https://www.statista.com/statistics/278249/global-

number-of-etfs/

Statista Research Department. (2024). Worldwide ETF

assets under management 2003-2023. Statista.

Retrieved from

https://www.statista.com/statistics/224579/worldwide-

etf-assets-under-management-since-1997/

Wang, J., & Zhang, G. (2024). Predicting relative

performance of sector exchange-traded funds using

machine learning. Journal of Business & Behavioral

Sciences, 36(1).

Yüksel, B. (2023). Comparative analysis of LSTM model

in predicting ETF stock prices for different sectors

(Master's thesis, Marmara Universitesi, Turkey).

ETF Forecast: Application of Innovative Machine Learning Models in the Field of ETF Forecasting

501