GaussianMixtureHMM

Here, we show model performance in predicting the bitcoin price and their hyper-parameters.

Prediction

Prices

DATE	CLOSE
2024-11-11	56,651.54
2024-11-18	56,179.75
2024-11-25	56,952.46
2024-12-02	56,559.88

Charts

Model Performance

For scoring, Root Mean Squared Error was used.

Training data

Mean across all targets: 4,974.43

y_CLOSE_7	y_CLOSE_14	y_CLOSE_21	y_CLOSE_28
2,765.86	2,321.75	6,985.47	7,824.65

Test data

Mean across all targets: 21,662.08

y_CLOSE_7	y_CLOSE_14	y_CLOSE_21	y_CLOSE_28
19,145.45	21,541.97	24,262.98	21,697.92

Model Setup

Below are choices made for the model setup.

Exogenous Data

(Determined by Forward Selection and heuristically based on the model's assumptions)

param	description	value
include_indices	Whether to include market indices.	False
include_commodities	Whether to include commodities.	False
include_currencies	Whether to include currencies.	True

Feature Generator

(Determined by Grid Search)

param	description	value
ohlc_ratio	Whether to add OHLC ratios of bitcoin price to the data.	True

Feature Scaler

(Determined by heuristically, considering the model's assumptions)

param	description	value
log	Whether to take the log of the data.	True
diff	Whether to difference the data.	True
yeo_johnson	Whether to apply Yeo-Johnson transformation to the data.	False
normalize	Whether to normalize the data.	True

Hyper-parameters

(Determined by Grid Search)

param	description	value
n_mix	The number of mixtures in each state.	5
n_components	The number of states.	3
n_train_iter	The number of training iterations.	100
n_auto_regress_iter	The number of iterations to run simulations to generate predictions. The averages of the values are used.	30
window_size	The size of the window to use. Overlapping windows are used.	5