| |
| author |
Gabriel Dutra
| | title |
Forecasting Financial Time Series: An Empirical Analysis of LSTM Model Performance Across Minute, Hourly, and Daily Time Frames
| | abstract |
Time series forecasting plays a crucial role in the world of finance. Whether in calculating risk
management, or taking investment decisions, the ability to estimate future trends accurately is vital to
maximizing profit. In recent years, machine learning algorithms gained popularity for being able to
obtain better results in forecasting than previous statistical models. The success of these algorithms,
however, relies heavily on selecting the correct features of the data, as well as fine-tuning multiple
parameters in the machine learning model. Although research in this field is extensive, most published
articles use daily historical financial data to perform analysis and draw results, while also relying on large
training sets. Experiments that use small training sets, or different time frames, such as hourly and
minute data, are still uncommon, which leaves the question of whether results obtained with a single
time frame and training set can be generalized. To answer these questions, we developed a framework
that creates, trains, and evaluates Long Short Term Memory (LSTM) neural network models for minute,
hourly, and daily data of a given time series. After testing the framework with 2022 and 2023 data from
the SPDR SP 500 ETF (SPY) and NASDAQ 100 (NQ) quotes, as well as 2020 and 2021 Bitcoin (BTC) data,
we found that utilizing hourly data improved the forecasting results of our model when compared to the
daily data. We also found that models trained on minute data failed to capture any trend in the data,
resulting in higher forecasting errors when compared to the other two time frames.
| | school |
The College of Liberal Arts, Drew University
| | degree |
B.S. (2023)
|
| advisor |
Dr. Steven Kass
|
| committee |
Dr. Hamed Yousefi
Dr. Yi Lu
|
| full text | GDutra.pdf |
| |
