3.3 Explanation and Implications
Table 1 summarizes the performance of the four
models concerning for the Dogecoin 1-hour interval
price. As shown, The LSTM model performed the
best with an MSE of only 0.0000035197, followed
closely by the linear regression model
(0.0000035667). The random forest model is worse in
terms of mean squared error performance
(0.0000040188), while the Transformer model
performs the worst, with an MSE value of
0.0000180260. The result shows that the LSTM
model is particularly good at capturing temporal
dependencies in sequential data., which is crucial for
time series prediction tasks. The LSTM model was
designed to remember long-term dependencies while
efficiently ignoring irrelevant information, making
them well-suited for predicting short-term trends and
fluctuations in DOGE prices.
LSTM has several advantages over linear
regression and random forest in terms of time series
forecasting. First and foremost, Time series data
including both short- and long-term dependencies
may be precisely captured using LSTM for it has the
ability to store data from earlier time steps and utilize
it to make predictions in the future. In contrast, linear
regression only captures linear relationships between
variables and cannot handle complex temporal
dependencies. Random forest, while capable of
handling nonlinear relationships, is based on decision
trees and is not inherently designed to process
sequential data with time dependencies. Meanwhile,
LSTM can naturally handle nonlinear relationships in
time series. Last but not least, the unique feature of
LSTM is its built-in memory units and forget gates,
which can selectively remember or forget information,
effectively filtering out noise and retaining useful
information. This mechanism allows LSTM to excel
in handling long sequences. Linear Regression and
Random Forest do not have this memory and
selective forgetting mechanism, so their ability to
capture patterns in long time series data is limited.
The transformer model, however, while highly
effective in tasks like natural language processing,
may not be as naturally suited to the time series
prediction task here. Transformers rely on self-
attention mechanisms, which, although excellent at
handling long-range dependencies, may struggle with
capturing the short-term dynamics typical in time
series data. Moreover, the dataset of Dogecoin is
relatively small, the Transformer model can’t learn
the patterns in the time series effectively, leading to
poorer performance.
This experiment explored the accuracy of four
models in predicting the price of Dogecoin and found
a better way to predict the price of Dogecoin. To go
further, some new laws or patterns discovered in the
experiment can become the basis for future research
on time series problems or provide new research
directions for such tasks. At the same time, for
investors, this experiment can help them better
understand the virtual currency represented by
Dogecoin and make profits from investment. Thus,
they can better understand market behavior and
market rules in complex changes, and then explore
how to use predictive models for investing and risk
aversion.
3.4 Limitations and Prospects
This study still has certain limitations. In terms of
data, the hourly price frequency of Dogecoin is too
low, and the price every 5 seconds may help the
model predict more accurately. At the same time, the
number of features of Dogecoin is too small, and
higher-dimensional data can be further collected. In
terms of models, due to limitations of experimental
conditions, the author cannot explore all models, but
there are indeed many valuable models that can be
explored in terms of price prediction of Dogecoin. To
improve this research, the author intends to examine
more methods, such as CNN or the Diffusion model.
At the same time, the author plans to collect more
frequent and more dimensional price data for further
research.
4 CONCLUSIONS
To sum up, this study investigates machine learning
techniques based on sample characteristics of
samples and dimensions to predict Dogecoin prices.
This paper compares the performance of four models,
i.e., Linear Regression, Random Forest, LSTM, and
Transformer, in predicting Dogecoin prices, and finds
that the LSTM model delivers the best results. The
results show that the LSTM model performs much
better than the others because of its robustness in
handling nonlinear interactions and capturing
temporal dependencies. Linear Regression and
Random Forest models underperformed in dealing
with complex time series data, while the Transformer
model also did not meet expectations in this task. The
limitation of this article is that the data volume and
feature dimensions of the sample are small, which
affects the prediction effect of the model. Future
research could focus on optimizing hyperparameters