Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Forecasting Bitcoin Price Movements Using Neural Networks: A Time Series Approach

Naujokas, Julius LU (2025) In Bachelor’s Theses in Mathematical Sciences MASK11 20251
Mathematical Statistics
Abstract
The extreme volatility and pronounced non-stationarity of cryptocurrency prices challenge traditional time-series models and motivate
the use of deep learning. This thesis addresses three focused research
questions: (i) Do gated recurrent neural networks (GRNNs) produce
more accurate one-day-ahead Bitcoin return predictions than feedforward neural networks and linear benchmarks? (ii) Can a GRNN
based forecasting model generate risk-adjusted return predictions to
be used in a fee-adjusted trading strategy? and (iii) What sampling
frequency is most suitable for training GRNN-based Bitcoin returnforecasting models? Minute level Bitcoin prices from 2018 – 2025 are
compressed to daily closes and combined with a small set of technical... (More)
The extreme volatility and pronounced non-stationarity of cryptocurrency prices challenge traditional time-series models and motivate
the use of deep learning. This thesis addresses three focused research
questions: (i) Do gated recurrent neural networks (GRNNs) produce
more accurate one-day-ahead Bitcoin return predictions than feedforward neural networks and linear benchmarks? (ii) Can a GRNN
based forecasting model generate risk-adjusted return predictions to
be used in a fee-adjusted trading strategy? and (iii) What sampling
frequency is most suitable for training GRNN-based Bitcoin returnforecasting models? Minute level Bitcoin prices from 2018 – 2025 are
compressed to daily closes and combined with a small set of technical indicator features. After cross-validated tuning, a Long ShortTerm Memory (LSTM) network dominates the Multi-Layer Perceptron (MLP) across all performance metrics. To turn point predictions
into usable risk metrics, we complete Monte Carlo simulations. While
the cubic OLS achieved marginally higher directional accuracy in onestep-ahead forecasts, its narrow prediction intervals made it unreliable
under regime shifts. The LSTM consistently overestimated future returns. In short, OLS wins the immediate point-forecast contest, but
neither perform well in an application scenario. Our experiments show
that, for this GRNN trading framework, daily data offers the optimal balance between signal quality and noise, making it the preferred
resolution over minute level data. (Less)
Popular Abstract
Cryptocurrencies like Bitcoin are famously unpredictable, making it difficult to build models that forecast prices accurately. In this thesis, I compare traditional statistics models with modern deep learning techniques to see which perform better at daily bitcoin return prediction. I particularly focus on Long Short-Term Memory (LSTM) models. While a classic regression model with a non-linear term (cubic OLS) fractionally outperform the LSTM in short-term accuracy, it failed to adapt when market conditions changed. The LSTM, on the other hand, offered more flexible forecasts, but often expected too much growth. Neither model would be ready for use in real world trading. The experiments show that using daily data – rather than high... (More)
Cryptocurrencies like Bitcoin are famously unpredictable, making it difficult to build models that forecast prices accurately. In this thesis, I compare traditional statistics models with modern deep learning techniques to see which perform better at daily bitcoin return prediction. I particularly focus on Long Short-Term Memory (LSTM) models. While a classic regression model with a non-linear term (cubic OLS) fractionally outperform the LSTM in short-term accuracy, it failed to adapt when market conditions changed. The LSTM, on the other hand, offered more flexible forecasts, but often expected too much growth. Neither model would be ready for use in real world trading. The experiments show that using daily data – rather than high frequency minute data – offers the most reliable balance between signal quality and noise, making it the preferred sampling frequency. (Less)
Please use this url to cite or link to this publication:
author
Naujokas, Julius LU
supervisor
organization
course
MASK11 20251
year
type
M2 - Bachelor Degree
subject
keywords
Bitcoin forecasting, Time series predictions, Non-linear regression, Monte Carlo simulation, LSTM
publication/series
Bachelor’s Theses in Mathematical Sciences
report number
LUNFMS-4081-2025
ISSN
1654-6229
other publication id
2025:K10
language
English
id
9209903
date added to LUP
2025-08-18 11:18:20
date last changed
2025-08-18 11:18:20
@misc{9209903,
  abstract     = {{The extreme volatility and pronounced non-stationarity of cryptocurrency prices challenge traditional time-series models and motivate
the use of deep learning. This thesis addresses three focused research
questions: (i) Do gated recurrent neural networks (GRNNs) produce
more accurate one-day-ahead Bitcoin return predictions than feedforward neural networks and linear benchmarks? (ii) Can a GRNN
based forecasting model generate risk-adjusted return predictions to
be used in a fee-adjusted trading strategy? and (iii) What sampling
frequency is most suitable for training GRNN-based Bitcoin returnforecasting models? Minute level Bitcoin prices from 2018 – 2025 are
compressed to daily closes and combined with a small set of technical indicator features. After cross-validated tuning, a Long ShortTerm Memory (LSTM) network dominates the Multi-Layer Perceptron (MLP) across all performance metrics. To turn point predictions
into usable risk metrics, we complete Monte Carlo simulations. While
the cubic OLS achieved marginally higher directional accuracy in onestep-ahead forecasts, its narrow prediction intervals made it unreliable
under regime shifts. The LSTM consistently overestimated future returns. In short, OLS wins the immediate point-forecast contest, but
neither perform well in an application scenario. Our experiments show
that, for this GRNN trading framework, daily data offers the optimal balance between signal quality and noise, making it the preferred
resolution over minute level data.}},
  author       = {{Naujokas, Julius}},
  issn         = {{1654-6229}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Bachelor’s Theses in Mathematical Sciences}},
  title        = {{Forecasting Bitcoin Price Movements Using Neural Networks: A Time Series Approach}},
  year         = {{2025}},
}