In theory, CNN works quite well in capturing
patterns; however, its performance in the present
study is not very impressive as clearly deduced from
Fig. 5. To be specific, the coefficient of determination
(R²) of the CNN model is -0.0369, which is dwarfed
by the OLS and XGBoost models, most likely
because the model has been able to capture very little
useful information. What this demonstrates is that the
OLS outperforms R² of these models where R²
approaches the range of zero or thereabouts are rare.
This shows the limitation of this CNN in such,
Keywords: R², degeneracy, prediction accuracy,
convolution neural networks.
In addition, it should be mentioned that the
performance of the CNN model is disappointing
when measured by MSE or MAE metrics. The MSE
was recorded at 0.000682, almost four times the
shagger OLS model, which implied an even larger
difference between the predicted and the actual
outcome. The MAE is 0.0179, which as well fixed the
high value and cannot be related to OLS models and
XGBoost models. Such high error indicators imply
that the CNN model is not very reliable in making
predictions of Bitcoin per day yield and predictions
made have fairly conspicuous bias.
In light of these results, it can be inferred that the
CNN model could be deficient for many reasons. First,
even though CNN is well-established and enjoyed
great advances in image processing, its convolution
operation may not be sufficient for forecasting, and
capturing the complex dependencies of time-series
financial data. More particularly, with regards to
financial assets like Bitcoin which tend to be very
volatile as well as nonlinear in nature, learning useful
patterns from sparse training data can be a daunting
task for CNN. Second, the model design and
parameterization of the CNN may also be prohibitive.
3.3 Discussion and Recommendations
The OLS model is satisfactory, evidenced by R² value
of 0.7499 meaning the model is capable of explaining
around 75% of the variability of such data.
Additionally, both MSE and MAE's mean square
error recording low figures of 0.000164 and 0.0076
respectively. This preliminary finding serves to
confirm that the daily return on Bitcoin can be
modelled with a certain degree of stability linearity
which the OLS can model more correctly when it is
not in a volatile market.
However, in complex and volatile environments,
additional knowledge is rendered from the XGBoost
model. It indicates that the R-squared measure
(0.52891) among considered XGBoost approaches is
not especially impressive, although it does indicate
some potential in a nonlinear modeling context. Due
to the ensemble approach employed by XGBoost,
various decision trees that the algorithm contains,
XGBoost is more adaptive to abrupt changes in the
market. Despite the fact that XGBoost has an MSE of
0.00031 and a MAE of 0.0123, both higher than the
OLS model, such errors seem to be typical for such
types of models, 'XGBoost' may be useful in targeting
reversals and abnormal returns.
On the other hand, the accuracy of the CNN model
leaves much to be desired. The R-squared figure for
the CNN model can be calculated at -0.0369, where
all other external models that have OLS and XGBoost
have fared better, which means that CNN has almost
closed out on viable market data. Its MSE is 0.000682
and MAE is 0.0179, which are not promising results
compared with any reasonable prediction in this task.
From the insights provided through the above
discussion, this study obtained several key
investment suggestions. First, in a less favorable
market, passive investments using statistical linear
models based on OLS may be efficient and would
stabilize returns, which is quite useful for low-risk
individuals. Second, when volatile, an investor can
seek complex relations based on nonlinear models
such as XGBoost, though this has its limitations as
there are large prediction errors that come with such
models. However, while parts of deep learning
models like CNN are efficient in other areas like
image classification, in the treatment of daily yield of
Bitcoin, where there is financial time series data is
high volatility and complex, their merits do not shine
as they should, thus a combination with other
approaches focusing on reducing errors is required.
3.4 Limitations
However, this study has some shortcomings and
defetcs. First, the dataset do not include certain
phenomena for which many experts seek correlation
with the price of bitcoin, such as macroeconomic
variables or voting behavior of the market; secondly,
hyper-parameter tuning of XGBoost and CNN is also
not sufficiently addressed. This could be a limitation
on the maximum ability of these models; furthermore,
this study has not examined the use of combined
models and their predictive enhancement. Future
studies in this area may be directed to the broadening
intensity of the data set scope, further focusing on the
model parameters, employing less rigid data layer
segmentation strategies, and assessing the potential of
hybrid models in enhancing the prediction effect.