Final Project

Name: M .Shaban ,Fahad ,Hamza Tahir ,Ahmed Saeed

Section: Fin-Tech B

Instructor: Ma’am Falak Khan

 Heteroskedasticity

Null hypothesis 𝐻0 : The error variance is constant

Alternative hypothesis 𝐻1 : The error variance is not constant

Breusch-Pagan-Godfrey
Heteroskedasticity Test: Breusch-Pagan-Godfrey
Null hypothesis: Homoskedasticity

F-statistic 17.40793 Prob. F(5,144) 0.0000

Obs*R-squared 56.50955 Prob. Chi-Square(5) 0.0000
Scaled explained SS 71.58488 Prob. Chi-Square(5) 0.0000

Test Equation:
Dependent Variable: RESID^2
Method: Least Squares
Date: 12/04/25 Time: 01:18
Sample: 2013M04 2025M09
Included observations: 150

Variable Coefficient Std. Error t-Statistic Prob.

C 2113671. 340430.7 6.208813 0.0000

GOLD -2119.659 410.9944 -5.157392 0.0000
SILVER 78823.07 15946.92 4.942841 0.0000
VOLUME 3.85E-07 1.24E-07 3.111426 0.0022
CRUDE_OIL -8802.995 2076.575 -4.239190 0.0000
MARKET_CAP 8.15E-07 1.78E-07 4.583331 0.0000

R-squared 0.376730 Mean dependent var 335708.4

Adjusted R-squared 0.355089 S.D. dependent var 558480.8
S.E. of regression 448495.5 Akaike info criterion 28.90436
Sum squared resid 2.90E+13 Schwarz criterion 29.02479
Log likelihood -2161.827 Hannan-Quinn criter. 28.95329
F-statistic 17.40793 Durbin-Watson stat 1.204003
Prob(F-statistic) 0.000000

Interpretation:A Breusch-Pagan-Godfrey heteroskedasticity test rejects the null of

homoskedasticity (Obs*R-squared = 56.51, p-value = 0.0000; F-statistic = 17.41, p-value =
0.0000), indicating significant heteroskedasticity in the residuals.
Harvey test output:

Null hypothesis 𝐻0 : The error variance is constant

Alternative hypothesis 𝐻1 : The error variance is not constant

Heteroskedasticity Test: Harvey

Null hypothesis: Homoskedasticity

F-statistic 11.63958 Prob. F(5,144) 0.0000

Obs*R-squared 43.17396 Prob. Chi-Square(5) 0.0000
Scaled explained SS 54.72866 Prob. Chi-Square(5) 0.0000

Test Equation:
Dependent Variable: LRESID2
Method: Least Squares
Date: 12/04/25 Time: 01:23
Sample: 2013M04 2025M09
Included observations: 150

Variable Coefficient Std. Error t-Statistic Prob.

C 15.38947 1.635156 9.411620 0.0000

GOLD -0.002416 0.001974 -1.224055 0.2229
SILVER 0.118367 0.076596 1.545337 0.1245
VOLUME 3.80E-13 5.95E-13 0.639538 0.5235
CRUDE_OIL -0.053447 0.009974 -5.358481 0.0000
MARKET_CAP 1.67E-12 8.54E-13 1.955818 0.0524

R-squared 0.287826 Mean dependent var 11.14218

Adjusted R-squared 0.263098 S.D. dependent var 2.509480
S.E. of regression 2.154213 Akaike info criterion 4.411906
Sum squared resid 668.2512 Schwarz criterion 4.532331
Log likelihood -324.8929 Hannan-Quinn criter. 4.460831
F-statistic 11.63958 Durbin-Watson stat 1.073744
Prob(F-statistic) 0.000000

Interpretation:The Harvey heteroskedasticity test strongly rejects the null of

homoskedastic errors (F-statistic = 11.64, p-value = 0.0000; Obs*R-squared = 43.17, p-value =
0.0000), indicating that the error variance is not constant and the model suffers from
heteroskedasticity
 Null hypothesis 𝐻0 : The error variance is constant
Alternative hypothesis 𝐻1 : The error variance is not constant

Heteroskedasticity Test: Glejser

Null hypothesis: Homoskedasticity

F-statistic 19.91646 Prob. F(5,144) 0.0000

Obs*R-squared 61.32361 Prob. Chi-Square(5) 0.0000
Scaled explained SS 69.31814 Prob. Chi-Square(5) 0.0000

Test Equation:
Dependent Variable: ARESID
Method: Least Squares
Date: 12/04/25 Time: 01:33
Sample: 2013M04 2025M09
Included observations: 150

Variable Coefficient Std. Error t-Statistic Prob.

C 1581.020 225.7507 7.003387 0.0000

GOLD -1.180552 0.272544 -4.331607 0.0000
SILVER 46.15801 10.57492 4.364857 0.0000
VOLUME 1.93E-10 8.21E-11 2.346480 0.0203
CRUDE_OIL -7.860679 1.377044 -5.708370 0.0000
MARKET_CAP 5.05E-10 1.18E-10 4.286896 0.0000

R-squared 0.408824 Mean dependent var 438.2576

Adjusted R-squared 0.388297 S.D. dependent var 380.2667
S.E. of regression 297.4120 Akaike info criterion 14.26729
Sum squared resid 12737362 Schwarz criterion 14.38772
Log likelihood -1064.047 Hannan-Quinn criter. 14.31622
F-statistic 19.91646 Durbin-Watson stat 0.887942
Prob(F-statistic) 0.000000

Interpretation:The Glejser heteroskedasticity test rejects the null of homoskedasticity

(F-statistic ≈ 19.92, p-value = 0.0000; Obs*R-squared ≈ 61.32, p-value = 0.0000), confirming
the presence of heteroskedasticity in the residuals
Solve by using log

Heteroskedasticity Test: Breusch-Pagan-Godfrey

Null hypothesis: Homoskedasticity

F-statistic 2.885598 Prob. F(5,144) 0.0163

Obs*R-squared 13.66045 Prob. Chi-Square(5) 0.0179
Scaled explained SS 26.36368 Prob. Chi-Square(5) 0.0001

Test Equation:
Dependent Variable: RESID^2
Method: Least Squares
Date: 12/04/25 Time: 01:42
Sample: 2013M04 2025M09
Included observations: 150

Variable Coefficient Std. Error t-Statistic Prob.

C -0.002580 0.015931 -0.161936 0.8716

LOG_GOLD 0.004245 0.003738 1.135743 0.2580
LOG_SILVER 0.001269 0.002935 0.432258 0.6662
LOG_VOLUME 0.000217 9.15E-05 2.366057 0.0193
LOG_MARKET_CAP -0.001166 0.000346 -3.371699 0.0010
LOG_OIL -0.001471 0.000885 -1.662160 0.0987

R-squared 0.091070 Mean dependent var 0.001472

Adjusted R-squared 0.059510 S.D. dependent var 0.003022
S.E. of regression 0.002931 Akaike info criterion -8.788076
Sum squared resid 0.001237 Schwarz criterion -8.667651
Log likelihood 665.1057 Hannan-Quinn criter. -8.739151
F-statistic 2.885598 Durbin-Watson stat 0.257030
Prob(F-statistic) 0.016313

Interpretation:After re-estimating the model with log-transformed regressors, the

Breusch-Pagan-Godfrey test still rejects the null of homoskedasticity (F-statistic = 2.89, p-value
= 0.016; Obs*R-squared = 13.66, p-value = 0.018). This indicates that some remaining
heteroskedasticity is present, although the problem is weaker than in the level-specification
HAC TEST
Dependent Variable: LOG_BITCOIN_PRICE
Method: Least Squares
Date: 12/04/25 Time: 02:05
Sample: 2013M04 2025M09
Included observations: 150
HAC standard errors & covariance (Bartlett kernel, Newey-West fixed
bandwidth = 5.0000)

Variable Coefficient Std. Error t-Statistic Prob.

C -14.80723 0.451878 -32.76822 0.0000

LOG_GOLD -0.192594 0.117120 -1.644413 0.1023
LOG_SILVER 0.236471 0.091338 2.588950 0.0106
LOG_VOLUME -0.006660 0.002212 -3.010276 0.0031
LOG_MARKET_CAP 0.959814 0.007511 127.7932 0.0000
LOG_OIL 0.007036 0.025635 0.274485 0.7841

R-squared 0.999657 Mean dependent var 8.644001

Adjusted R-squared 0.999645 S.D. dependent var 2.078545
S.E. of regression 0.039153 Akaike info criterion -3.603493
Sum squared resid 0.220748 Schwarz criterion -3.483068
Log likelihood 276.2620 Hannan-Quinn criter. -3.554568
F-statistic 83956.14 Durbin-Watson stat 0.206091
Prob(F-statistic) 0.000000 Wald F-statistic 38626.36
Prob(Wald F-statistic) 0.000000

Interpretation:Due to evidence of heteroskedasticity and autocorrelation, the final log–log

Bitcoin price model is estimated using HAC (Newey–West) robust standard errors. Under HAC,
silver and market capitalization remain positively and significantly related to Bitcoin price, while
trading volume has a small but significant negative effect. In contrast, gold and oil prices are not
statistically significant once robust standard errors are used.
 Autocorrelation

Null hypothesis 𝐻0 : There is no serial correlation

Alternative hypothesis 𝐻1 : There is serial correlation

Interpretation:The Breusch–Godfrey LM test (lag 2) strongly rejects the null of no

serial correlation (Obs*R-squared = 123.81, p-value = 0.0000), indicating substantial
 autocorrelation in the residuals. Therefore, the final model is estimated using HAC (Newey–
West) robust standard errors, which are consistent in the presence of both heteroskedasticity
and autocorrelation

Normality test

Null hypothesis 𝑯𝟎 : The residuals are normally distributed

Alternative hypothesis 𝑯𝟏 : The residuals are not normally distributed

Interpretation:The residual histogram indicates a non-normal distribution, with

pronounced positive skewness (1.52) and excess kurtosis (5.19). The Jarque–Bera statistic of
87.92 (p-value = 0.0000) strongly rejects the null hypothesis of normal residuals

Multicollinearity
 Null hypothesis 𝐻0 : There is no serious multicollinearity
Alternative hypothesis 𝐻1 : There is serious multicollinearity

Interpretation:Pairwise correlation coefficients among the explanatory variables are

generally below 0.9. The highest correlation is between log volume and log market
capitalization (about 0.88), indicating some degree of association but not perfect collinearity.
Overall, the correlation matrix does not suggest severe multicollinearity.

ADF
Null hypothesis 𝑯𝟎 : this variables has a unit root (non-stationary)
Alternative hypothesis 𝑯𝟏 : this variables has no a unit root (stationary)
Null Hypothesis: LOG_MARKET_CAP has a unit root
Exogenous: None
Lag Length: 0 (Automatic - based on SIC, maxlag=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic 2.324621 0.9953

Test critical values: 1% level -2.580574
5% level -1.942982
10% level -1.615289

*MacKinnon (1996) one-sided p-values.

Null Hypothesis: LOG_OIL has a unit root

Exogenous: None
Lag Length: 2 (Automatic - based on SIC, maxlag=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -0.435383 0.5243

Test critical values: 1% level -2.580788
5% level -1.943012
10% level -1.615270

*MacKinnon (1996) one-sided p-values.

Null Hypothesis: LOG_SILVER has a unit root
Exogenous: None
Lag Length: 1 (Automatic - based on SIC, maxlag=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -0.101253 0.6472

Test critical values: 1% level -2.580681
5% level -1.942996
10% level -1.615279

*MacKinnon (1996) one-sided p-values.

Null Hypothesis: LOG_VOLUME has a unit root
Exogenous: None
Lag Length: 0 (Automatic - based on SIC, maxlag=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic 0.558960 0.8359

Test critical values: 1% level -2.580574
5% level -1.942982
10% level -1.615289

*MacKinnon (1996) one-sided p-values.

Interpretation:Because the ADF statistic is greater than all critical values and the p-value
is much bigger than 0.05, you cannot reject 𝐻0 . So all variables is non-stationary in levels; it
behaves like an integrated process.

Now first differences

Null Hypothesis: D(LOG_BITCOIN_PRICE) has a unit root
Exogenous: None
Lag Length: 0 (Automatic - based on SIC, maxlag=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -10.86769 0.0000

Test critical values: 1% level -2.580681
5% level -1.942996
10% level -1.615279
Null Hypothesis: D(LOG_SILVER) has a unit root
Exogenous: None
Lag Length: 0 (Automatic - based on SIC, maxlag=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -9.432512 0.0000

Test critical values: 1% level -2.580681
5% level -1.942996
10% level -1.615279

*MacKinnon (1996) one-sided p-values.

Null Hypothesis: D(LOG_MARKET_CAP) has a unit root

Exogenous: None
Lag Length: 0 (Automatic - based on SIC, maxlag=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -10.79859 0.0000

Test critical values: 1% level -2.580681
5% level -1.942996
10% level -1.615279

*MacKinnon (1996) one-sided p-values.

 Null Hypothesis: D(LOG_OIL) has a unit root
Exogenous: None
Lag Length: 1 (Automatic - based on SIC, maxlag=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -9.132264 0.0000

Test critical values: 1% level -2.580788
5% level -1.943012
10% level -1.615270

*MacKinnon (1996) one-sided p-values.

Null Hypothesis: D(LOG_VOLUME) has a unit root

Exogenous: None
Lag Length: 0 (Automatic - based on SIC, maxlag=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -12.27495 0.0000

Test critical values: 1% level -2.580681
5% level -1.942996
10% level -1.615279

*MacKinnon (1996) one-sided p-values.

Interpretation:ADF tests on the first differences of the logged series

(D(LOG_BITCOIN_PRICE), D(LOG_GOLD), D(LOG_SILVER), D(LOG_OIL),
D(LOG_VOLUME), D(LOG_MARKET_CAP)) strongly reject the unit root hypothesis at the
1% significance level (test statistics far below the 1% critical values and p-values equal to
0.0000), implying that all series are stationary in first differences and can be treated as integrated
of order one, I(1).

KPSS tests
Null hypothesis 𝑯𝟎 : The series is stationary
Alternative hypothesis 𝑯𝟏 : The series is non-stationary
Null Hypothesis: LOG_BITCOIN_PRICE is stationary
Exogenous: Constant
Bandwidth: 10 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 1.388276

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 4.291548

HAC corrected variance (Bartlett kernel) 42.66018

Null Hypothesis: LOG_GOLD is stationary

Exogenous: Constant
Bandwidth: 10 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 1.047294

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 0.049251

HAC corrected variance (Bartlett kernel) 0.473685

Null Hypothesis: LOG_MARKET_CAP is stationary

Exogenous: Constant
Bandwidth: 10 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 1.395900

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 4.911446

HAC corrected variance (Bartlett kernel) 48.89081
Null Hypothesis: LOG_VOLUME is stationary
Exogenous: Constant
Bandwidth: 9 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 1.354330

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 44.99407

HAC corrected variance (Bartlett kernel) 357.6220

Interpretation:Because test statistic is much larger than all critical values, you reject 𝐻0 at
the 1% level and conclude that LOG_BITCOIN_PRICE , LOG_GOLD, LOG_VOLUME,
LOG_MARKET_CAP is not stationary. This matches your ADF result (which also said
non-stationary in levels)

Null Hypothesis: LOG_OIL is stationary

Exogenous: Constant
Bandwidth: 9 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 0.235569

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 0.114700

HAC corrected variance (Bartlett kernel) 0.930390

Interpretation:Because LM-stat is below all critical values, you do not reject 𝐻0, so
LOG_OIL is treated as stationary in level according to KPSS.
Null Hypothesis: LOG_SILVER is stationary
Exogenous: Constant
Bandwidth: 9 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 0.443314

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 0.053628

HAC corrected variance (Bartlett kernel) 0.428263

Interpretation: Because LM-stat is below all critical values, you do not reject 𝐻0, so
LOG_Silver is treated as stationary in level according to KPSS.

Now first differences

Null Hypothesis: D(LOG_BITCOIN_PRICE) is stationary

Exogenous: Constant
Bandwidth: 0 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 0.063043

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 0.059372

HAC corrected variance (Bartlett kernel) 0.059372
Null Hypothesis: D(LOG_VOLUME) is stationary
Exogenous: Constant
Bandwidth: 4 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 0.299318

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 2.114704

HAC corrected variance (Bartlett kernel) 1.933116

Interpretation: Because LM-stat is below all critical values, you do not reject 𝐻0, so
LOG_Volume and Log_Bitcon_Price is treated as stationary in level according to KPSS.

Null Hypothesis: D(LOG_GOLD) is stationary

Exogenous: Constant
Bandwidth: 3 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 0.525448

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 0.001011

HAC corrected variance (Bartlett kernel) 0.001424

Interpretation: Because LM-stat is above all critical values, you reject 𝐻0 , so LOG_Gold
treated as stationary in level according to KPSS.
Null Hypothesis: D(LOG_MARKET_CAP) is stationary
Exogenous: Constant
Bandwidth: 0 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 0.077581

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 0.059428

HAC corrected variance (Bartlett kernel) 0.059428

Interpretation: Because LM-stat is below all critical values, you do not reject 𝐻0, so
LOG_Market_Cap is treated as stationary in level according to KPSS.

Now Sceond differences

Null Hypothesis: D(LOG_GOLD,2) is stationary
Exogenous: Constant
Bandwidth: 34 (Newey-West automatic) using Bartlett kernel

LM-Stat.

Kwiatkowski-Phillips-Schmidt-Shin test statistic 0.136254

Asymptotic critical values*: 1% level 0.739000
5% level 0.463000
10% level 0.347000

*Kwiatkowski-Phillips-Schmidt-Shin (1992, Table 1)

Residual variance (no correction) 0.001431

HAC corrected variance (Bartlett kernel) 6.01E-05

Interpretation: Because LM-stat is below all critical values, you do not reject 𝐻0, so
LOG_Gold is treated as stationary in level according to KPSS.
 ARIMA

Interpretation of ACF/PACF for Bitcoin Returns:

The ACF and PACF plots reveal a strong seasonal pattern at lags 12 and 24, indicating
monthly seasonality in Bitcoin returns. The significant spike at lag 1 in the PACF suggests a
non-seasonal AR(1) component, while the slow decay in seasonal lags points to a seasonal
AR(1) term at lag 12. Negative ACF spikes at lags 5–8 may indicate moving average effects.
These patterns support estimating a seasonal ARIMA (SARIMA) model with both non-
seasonal and seasonal autoregressive components to capture the periodic dynamics in
the data.
Estimate ARIMA(p,d,q)

Interpretation of ARIMA(1,0,1) Estimation:

The ARIMA(1,0,1) model shows that the MA(1) term is marginally significant (p=0.0667),
while the AR(1) term is insignificant (p=0.8250), suggesting that moving average effects
dominate in explaining Bitcoin returns volatility. The constant is highly significant,
indicating a positive average return. The model explains about 13.6% of the variation in
returns, and the Durbin-Watson statistic near 2 suggests no residual autocorrelation.
However, the high AIC (23.32) and the presence of seasonal spikes in earlier diagnostics
imply that a seasonal ARIMA (SARIMA) specification may better capture the data's
periodic structure.
Estimate SARIMA(p,d,q)(P,D,Q)s

Interpretation of SARIMA(1,0,1)×(1,0,1)₁₂ Estimation:

The SARIMA model shows a highly significant seasonal AR(12) term (p=0.0028),
confirming strong monthly seasonality in Bitcoin returns. While non-seasonal AR(1) and
MA(1) terms are insignificant, the inclusion of seasonal components improves model fit—
R-squared increases to 36.3% (vs. 13.6% in ARIMA). The lower AIC (23.07) suggests better
model performance, and the Durbin-Watson near 2 indicates well-behaved residuals. The
significant seasonal persistence supports the use of SARIMA for forecasting Bitcoin
returns, though non-seasonal dynamics may be better captured through seasonal terms
alone.

Model Estimation: VAR/VECM

Check Cointegration (Johansen Test)

Date: 12/04/25 Time: 01:31
Sample (adjusted): 3 150
Included observations: 148 after adjustments
Trend assumption: Linear deterministic trend
Series: BTC_RETURNS GOLD OIL SILVER
Lags interval (in first differences): 1 to 1
Interpretation of Johansen Cointegration Test
Unrestricted Cointegration Rank Test (Trace)
Results:
Hypothesized Trace 0.05
No. of CE(s) Eigenvalue Statistic Critical Value Prob.**

None *
At most 1 *
0.766334
0.659090
543.9648
328.7930
47.85613
29.79707
0.0000
0.0000
The Johansen cointegration test indicates 4
At most 2 *
At most 3 *
0.512731
0.347210
169.5249
63.12207
15.49471
3.841465
0.0000
0.0000 cointegrating relationships among Bitcoin returns,
Trace test indicates 4 cointegrating eqn(s) at the 0.05 level
* denotes rejection of the hypothesis at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values
gold, oil, and silver prices, as both Trace and Max-
Unrestricted Cointegration Rank Test (Maximum Eigenvalue) Eigenvalue statistics reject all null hypotheses at the 5%
Hypothesized
No. of CE(s) Eigenvalue
Max-Eigen
Statistic
0.05
Critical Value Prob.** level. This strong evidence of long-run equilibrium
None *
At most 1 *
At most 2 *
0.766334
0.659090
0.512731
215.1717
159.2681
106.4028
27.58434
21.13162
14.26460
0.0000
0.0000
0.0000
suggests that these series move together over time
At most 3 * 0.347210 63.12207 3.841465

Max-eigenvalue test indicates 4 cointegrating eqn(s) at the 0.05 level

0.0000
despite short-term fluctuations. The significant
* denotes rejection of the hypothesis at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values adjustment coefficients for Bitcoin returns and gold
Unrestricted Cointegrating Coefficients (normalized by b'*S11*b=I):

BTC_RETU... GOLD OIL SILVER

show they are responsive to deviations from
7.16E-05
-0.000106
-7.98E-05
-0.005001
0.013041
0.000343
-0.079366
-0.017641
-0.034399
0.341908
-0.534569
0.395743
equilibrium, with Bitcoin exhibiting the strongest error-
6.23E-05 -0.007114 0.003564 -0.006464
correction behavior. These findings support estimating
Unrestricted Adjustment Coefficients (alpha):
a Vector Error Correction Model (VECM) to capture
D(BTC_RET... -15336.35 19296.50 4364.601 2044.582
D(GOLD)
D(OIL)
-120.0179
14.76984
182.9498
5.194133
-7.924590
0.356505
87.41859
-1.270722 both long-run cointegrating relationships and short-run
D(SILVER) -0.188187 1.899219 -1.937682 0.396875

dynamics for further analysis and forecasting.

1 Cointegrating Equation(s): Log likelihood -3544.907

Normalized cointegrating coefficients (standard error in parentheses)

BTC_RETU... GOLD OIL SILVER
1.000000 -69.85788 -1108.550 4775.605
(5.09344) (57.0875) (442.889)

Adjustment coefficients (standard error in parentheses)

D(BTC_RET... -1.098001
(0.15735)
D(GOLD) -0.008593
(0.00167)
D(OIL) 0.001057
(6.4E-05)
D(SILVER) -1.35E-05
(2.2E-05)

2 Cointegrating Equation(s): Log likelihood -3465.273

Normalized cointegrating coefficients (standard error in parentheses)

BTC_RETU... GOLD OIL SILVER
1.000000 0.000000 -2790.132 4434.196
(147.448) (659.763)
0.000000 1.000000 -24.07146 -4.887197
(1.42941) (6.39598)

Adjustment coefficients (standard error in parentheses)

D(BTC_RET... -3.146967 328.3423
(0.19029) (20.7529)
D(GOLD) -0.028019 2.986041
(0.00225) (0.24540)
D(OIL) 0.000506 -0.006136
(9.9E-05) (0.01083)
D(SILVER) -0.000215 0.025708
(3.3E-05) (0.00364)

3 Cointegrating Equation(s): Log likelihood -3412.072

Normalized cointegrating coefficients (standard error in parentheses)

BTC_RETU... GOLD OIL SILVER
1.000000 0.000000 0.000000 -3992.239
(251.138)
0.000000 1.000000 0.000000 -77.58506
(2.53925)
0.000000 0.000000 1.000000 -3.020085
(0.16223)

Adjustment coefficients (standard error in parentheses)

D(BTC_RET... -3.495133 329.8390 726.6365
(0.21727) (20.1185) (127.126)
D(GOLD) -0.027387 2.983324 6.570491
(0.00265) (0.24530) (1.54999)
D(OIL) 0.000477 -0.006014 -1.276123
(0.00012) (0.01082) (0.06840)
D(SILVER) -6.06E-05 0.025044 0.048085
(3.1E-05) (0.00285) (0.01799)
Estimate VAR

Interpretation of VAR Model Estimation:

The VAR model demonstrates strong explanatory power, with high R-squared
values ranging from 83.7% (Bitcoin returns) to 91.2% (oil prices), indicating that the
included variables jointly capture a substantial portion of the variation in each series.
The model fits best for commodity prices (gold, oil, silver) compared to Bitcoin.
The AIC and SIC values suggest reasonable model fit, though the large number of
coefficients (111) may indicate over-parameterization. The system exhibits significant
interdependence, supporting further analysis through Granger causality tests and
impulse response functions to explore dynamic interactions among Bitcoin returns
and commodity markets.
Estimate GARCH(1,1)

Interpretation of GARCH(1,1) Model Estimation:

The GARCH(1,1) model shows evidence of volatility persistence (α + β = 0.7705), though

the sum is below the typical high-persistence threshold of 0.9. The negative ARCH term (-
0.0853) is unusual and statistically marginal, suggesting potential model
misspecification—volatility may not follow a standard GARCH process. The model did not
converge after 500 iterations, indicating estimation difficulties. The low and negative R-
squared implies poor fit for mean returns, while the significant GARCH term confirms time-
varying volatility. Alternative specifications (EGARCH, GJR-GARCH) or different error
distributions may better capture Bitcoin's volatility dynamics.
Estimate EGARCH(1,1)

EGARCH(1,1) Interpretation:

The EGARCH results appear identical to GARCH(1,1), suggesting incorrect

specification—asymmetric and log variance terms are missing. The model did not
converge, indicating estimation issues. Re-specify with proper EGARCH form to
assess volatility asymmetry and leverage effects in Bitcoin returns.

Estimate GJR-GARCH(1,1)
GJR-GARCH(1,1) Interpretation:

Same output as GARCH(1,1)—threshold term missing, suggesting incorrect

specification. No convergence after 500 iterations. Re-estimate with proper
asymmetry term to capture negative shock effects.

To test persistence > 0.9:

Interpretation of GARCH Persistence Test:

The estimated persistence (α + β = 0.6853) is below 0.9, indicating moderate but not
strong volatility persistence in Bitcoin returns. This suggests that volatility shocks
decay relatively quickly rather than persisting indefinitely, which is unusual for
financial assets that typically exhibit high persistence. This result may reflect Bitcoin's
unique market behavior or potential model misspecification, warranting further
investigation with alternative volatility models or distributions.

Residual Diagnostics
Interpretation of ARCH Residual Test:

The ARCH test on GARCH residuals yields a significant F-statistic (p=0.0480), indicating
remaining ARCH effects in the residuals. This suggests the GARCH(1,1) model does not
fully capture all volatility clustering, implying possible model inadequacy or the need for
higher-order ARCH/GARCH terms or alternative specifications.

Model selection (AIC, BIC)

Lag Order Selection

Null and Alternative Hypotheses:

H₀: The VAR model with fewer lags is sufficient (no significant improvement from adding
more lags).

H₁: Adding more lags significantly improves the model fit.

Interpretation

● The VAR lag selection table reports multiple information criteria (LR, FPE, AIC, SC,
HQ) used to determine the optimal number of lags for the model.

● The Likelihood Ratio (LR) test selects lag 6, indicating that adding lags up to 6
significantly improves model fit at the 5% level.
 ● The Final Prediction Error (FPE) criterion chooses lag 6, suggesting that a 6-lag
model provides the lowest prediction error among all tested lags.

● The Akaike Information Criterion (AIC) also selects lag 6, meaning this lag length
offers the best balance between model fit and complexity under AIC’s lesser penalty
for additional parameters.

● The Schwarz Criterion (SC/BIC) selects lag 1, because SC penalizes additional

parameters more heavily, favoring a more parsimonious model.

● The Hannan–Quinn (HQ) criterion also selects lag 1, consistent with its moderate
penalty for adding lags.

● Overall, three criteria (LR, FPE, AIC) favor lag 6, while two criteria (SC, HQ) favor lag
1, indicating a trade-off between model complexity and parsimony.

● In practice, lag 6 is appropriate when capturing richer dynamics is important, while

lag 1 is preferred when prioritizing simplicity and avoiding overfitting.

Granger Causality Tests

H₀: The other variables do not Granger-cause the dependent variable.

H₁: At least one of the other variables does Granger-cause the dependent variable.

Interpretation

● The table presents VAR Granger causality (Block Exogeneity Wald) test results for
each dependent variable to determine whether other variables help predict it.

● A variable Granger-causes another if the p-value < 0.05.

Dependent variable: BTC_RET

● All p-values are above 0.05.

● No variable Granger-causes BTC_RET, meaning none of the included variables

significantly improve the prediction of Bitcoin returns.

Dependent variable: VOL

● BTC_RET (p = 0.0207), MCAP (p = 0.0282), DLOG(GOLD) (p = 0.0031), and

DLOG(SILVER) (p = 0.0002) all have p < 0.05.

● BTC returns, market cap, gold log-returns, and silver log-returns Granger-cause VOL,
indicating they significantly predict changes in Bitcoin volatility.

● DLOG(OIL) does not Granger-cause VOL.

Dependent variable: MCAP

● All p-values > 0.05.

● None of the variables Granger-cause MCAP, meaning Bitcoin market capitalization

is not significantly explained by the others.

Dependent variable: DLOG(GOLD)

 ● All p-values > 0.05.

● No variable Granger-causes gold log-returns, indicating no predictive relationship

from BTC_RET, VOL, MCAP, SILVER, or OIL.

Dependent variable: DLOG(SILVER)

● All p-values > 0.05.

● No variable Granger-causes silver log-returns.

Dependent variable: DLOG(OIL)

● All p-values > 0.05.

● No variable Granger-causes oil log-returns.

Overall Summary

● Bitcoin returns (BTC_RET) are not Granger-caused by any variable.

● Bitcoin volatility (VOL) has the strongest causal relationships, being significantly
influenced by BTC_RET, MCAP, GOLD, and SILVER.

● Market cap, gold, silver, and oil variables do not appear to be Granger-caused by any
of the other variables in the system.

This indicates that volatility is the most responsive and interconnected variable in the
system, while the other variables behave more independently within the VAR framework.

Summary of Findings:

VOL is the only variable for which H₀ is rejected—meaning it is Granger-caused by other

variables in the system.

For all other variables (BTC_RET, MCAP, GOLD, SILVER, OIL), H₀ cannot be rejected—no
significant predictive relationships were found from the tested variables.
Impulse Response Analysis (IRF)
Forecasting and evaluation
Chow Breakpoint Test: 2020M03
Null Hypothesis: No breaks at specified breakpoints

Equation Sample: 2013M12 2025M09

F-statistic 1.283454 Prob. F(6,130) 0.2693

Log likelihood ratio 8.171844 Prob. Chi-Square(6) 0.2258
Dependent Variable: BTC_RET
Method: Least Squares
Date: 12/04/25 Time: 00:47
Sample: 2013M12 2020M02
Included observations: 75
HAC standard errors & covariance (Bartlett kernel, Newey-West fixed
bandwidth = 4.0000)

Variable Coefficient Std. Error t-Statistic Prob.

C -0.596030 1.034374 -0.576223 0.5663

GOLD_RET -1.590545 1.379529 -1.152962 0.2529
SILVER_RET 1.567922 0.878858 1.784044 0.0788
OIL_RET -0.119369 0.243377 -0.490468 0.6254
BTC_VOL_LN -0.003918 0.040669 -0.096330 0.9235
BTC_MCAP_LN 0.029717 0.075308 0.394603 0.6944

R-squared 0.060732 Mean dependent var 0.027067

Adjusted R-squared -0.007331 S.D. dependent var 0.219034
S.E. of regression 0.219836 Akaike info criterion -0.115255
Sum squared resid 3.334610 Schwarz criterion 0.070144
Log likelihood 10.32207 Hannan-Quinn criter. -0.041227
F-statistic 0.892293 Durbin-Watson stat 1.822305
Prob(F-statistic) 0.491218 Wald F-statistic 1.246641
Prob(Wald F-statistic) 0.297059

Dependent Variable: BTC_RET

Method: Least Squares
Date: 12/04/25 Time: 00:49
Sample: 2020M04 2025M09
Included observations: 66
HAC standard errors & covariance (Bartlett kernel, Newey-West fixed
bandwidth = 4.0000)

Variable Coefficient Std. Error t-Statistic Prob.

C -0.038991 0.661066 -0.058982 0.9532

GOLD_RET 2.860843 0.870017 3.288260 0.0017
SILVER_RET -1.608577 0.503173 -3.196870 0.0022
OIL_RET 0.154899 0.255743 0.605682 0.5470
BTC_VOL_LN -0.034491 0.028719 -1.200959 0.2345
BTC_MCAP_LN 0.037194 0.045562 0.816334 0.4175

R-squared 0.150808 Mean dependent var 0.043551

Adjusted R-squared 0.080042 S.D. dependent var 0.178246
S.E. of regression 0.170964 Akaike info criterion -0.608220
Sum squared resid 1.753720 Schwarz criterion -0.409160
Log likelihood 26.07126 Hannan-Quinn criter. -0.529562
F-statistic 2.131079 Durbin-Watson stat 1.627305
Prob(F-statistic) 0.073894 Wald F-statistic 4.722309
Prob(Wald F-statistic) 0.001056
1. The CUSUM Test (Cumulative Sum of Recursive Residuals)

• Purpose: This test detects systematic changes in the regression coefficients. It

checks if the relationship between your independent and dependent variables
has remained constant (stable) throughout the time period analyzed.

• The Components:

o The Blue Line (CUSUM): This represents the cumulative sum of recursive
residuals.

o The Red Dashed Lines (5% Significance): These act as the "Critical
Bounds" or confidence bands. They represent the threshold for
statistical significance at the 5% level.

General Interpretation Rules

• Null Hypothesis ($H_0$): The parameters (coefficients) are stable over time.

• Alternative Hypothesis ($H_1$): The parameters are unstable (i.e., there is a

structural break).

Interpretation of Your Specific Graph

To interpret your result, you look at the position of the blue line relative to the red
dashed lines:

• Observation: In your graph, the blue CUSUM line fluctuates but remains strictly
within the two red dashed lines (the critical bounds) for the entire duration of
the x-axis (observation periods 14 to 25).

• Conclusion: Because the blue line never crosses the red critical bounds, you
fail to reject the Null Hypothesis.

• Final Verdict: The model is structurally stable at the 5% significance level. This
means the coefficients in your econometric model have not changed
significantly over the period studied

.8
Forecast: BTC_RETF
Actual: BTC_RET
.6
Forecast sample: 2024M04 2025M09
Included observations: 18
.4
Root Mean Squared Error 0.137227
Mean Absolute Error 0.109262
.2 Mean Abs. Percent Error 170.2102
Theil Inequality Coef. 0.597591
.0 Bias Proportion 0.352021
Variance Proportion 0.480001
-.2 Covariance Proportion 0.167978
Theil U2 Coefficient 0.762635
-.4 Symmetric MAPE 112.4005
II III IV I II III
2024 2025
Image Probable Test Pass Condition (Stable Model)

1 CUSUM Blue line stays inside red lines.

2 CUSUM of Squares Blue line stays inside red lines.

Recursive Coeff Blue line is flat/horizontal; does not trend

3
(Variable 1) sharply.

Recursive Coeff Blue line is flat/horizontal; does not trend

4
(Variable 2) sharply.

Recursive Coeff Blue line is flat/horizontal; does not trend

5
(Variable 3) sharply.

Recursive Coeff Blue line is flat/horizontal; does not trend

6
(Variable 4) sharply.