Goal of this project

The main purpose of this project is to explore the aftermath of Elon Musk's tweet on May 13, 2022, where he announced that Tesla would no longer accept Bitcoin as a form of payment. This unexpected decision sent shockwaves through the Bitcoin community and had a noticeable impact on its price.

For that, I employed the powerful statistical package called Google Causal Impact to study this phenomenon. With its ability to estimate the causal impact of an event on a target variable while accounting for other influencing factors, it was the perfect tool for this analysis.

The results were truly enlightening! The analysis revealed a significant and negative causal effect of Musk’s tweet on Bitcoin’s price. It became clear that the tweet had triggered panic within the Bitcoin community and influenced investor sentiment, leading to a noticeable decrease in price.

Elon vs BTC

1 Introduction¶

1.1 Case Study¶

On May 13, 2022, Elon Musk, the CEO of Tesla and a prominent figure in the cryptocurrency community, made a significant announcement on Twitter. In his tweet, Musk stated that Tesla would no longer accept Bitcoin as a form of payment. This unexpected decision sent shockwaves through the Bitcoin community and had a noticeable impact on the price of Bitcoin.

To analyze the causal relationship between Musk's tweet and Bitcoin's price and what would be the alternative sceinario of the Bitcoin prices if Elon Musk did not make the tweet, we will employ the "Google Causal Impact" using Python.

1.2 Google Causal Impact¶

The "Google Causal Impact" is a statistical approach used to estimate the causal impact of a particular event on a target variable, while accounting for various other factors that may influence the outcome. It is based on Bayesian structural time series models and is particularly effective in evaluating the impact of stock/crypto price change, marketing campaigns, policy changes, or other interventions. Google Causal Impact package was initially developed for "R" by Google team, later the packege was imported to "python"

By using this technique, we can isolate the effect of Elon Musk's tweet on Bitcoin's price, controlling for other factors such as market trends, trading volume, and external events. The analysis will provide insights into the direct influence of the tweet on Bitcoin's price movement, allowing us to understand the magnitude and significance of the impact. And help to visualize the what-if scenario for the Bitcoin price.

2 Model Implementation¶

2.1 Libraries, Dates, Data¶

Import python libraries

Set the data time-frame & stocks

Import the stock datas

In [1]:

import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
import yfinance as yf
from causalimpact import CausalImpact

In [2]:

training_start = "2021-04-12"
training_end = "2021-05-12"
treatment_start = "2021-05-13"
treatment_end = "2021-05-20"
end_stock = "2021-05-21"

In [3]:

y_stock = ["BTC-USD"]
y = yf.download(y_stock,
                training_start,
                end_stock,
                interval="1d")
y.head()

[*********************100%***********************]  1 of 1 completed

Out[3]:

	Open	High	Low	Close	Adj Close	Volume
Date
2021-04-12	60175.945312	61253.035156	59589.875000	59893.453125	59893.453125	51828688519
2021-04-13	59890.019531	63742.285156	59869.957031	63503.457031	63503.457031	69983454362
2021-04-14	63523.753906	64863.097656	61554.796875	63109.695312	63109.695312	77451779687
2021-04-15	63075.195312	63821.671875	62208.964844	63314.011719	63314.011719	60954381579
2021-04-16	63258.503906	63594.722656	60222.531250	61572.789062	61572.789062	84293007468

In [4]:

y = y["Adj Close"].rename("BTC")
y.head(2)

Out[4]:

Date
2021-04-12    59893.453125
2021-04-13    63503.457031
Name: BTC, dtype: float64

In [5]:

X_stocks = ["MSFT", "NVDA", "AI", "EA", "SQ", "CRSP", "GOOG", "WMT"]
X = yf.download(X_stocks,
                training_start,
                end_stock,
                interval="1d")
X.head()

[*********************100%***********************]  8 of 8 completed

Out[5]:

	Adj Close								Close		...	Open		Volume
	AI	CRSP	EA	GOOG	MSFT	NVDA	SQ	WMT	AI	CRSP	...	SQ	WMT	AI	CRSP	EA	GOOG	MSFT	NVDA	SQ	WMT
Date
2021-04-12	59.889999	114.389999	139.680893	112.739502	250.827927	151.826202	265.200012	134.979538	59.889999	114.389999	...	258.190002	140.070007	3111900	1080900	2058100	31318000	27148700	86932400	9051300	6241000
2021-04-13	63.009998	120.879997	140.332626	113.363503	253.356720	156.523056	273.230011	134.564331	63.009998	120.879997	...	267.480011	139.800003	3297300	1352800	2105300	23310000	23837500	67621200	10561700	5799300
2021-04-14	68.459999	123.110001	139.838898	112.741997	250.514282	152.505051	258.399994	134.516068	68.459999	123.110001	...	274.130005	139.119995	11721300	1914000	1424900	20220000	23070900	38550000	12486400	7308200
2021-04-15	66.500000	123.849998	140.984390	114.833000	254.346634	161.092590	263.079987	135.327118	66.500000	123.849998	...	262.850006	139.399994	6223600	896300	1654300	27472000	25627500	59848000	8944400	7236800
2021-04-16	66.790001	118.559998	139.285889	114.888000	255.562057	158.849014	256.100006	135.761581	66.790001	118.559998	...	263.239990	140.889999	2988000	1359200	2468400	22596000	24878600	33520800	8549300	8829500

5 rows × 48 columns

💡 Here in confounders (X) stocks we are excluding stoks related to crypto currency & Elon Musk for remove selection biases

In [6]:

X = X["Adj Close"]
X.head(2)

Out[6]:

	AI	CRSP	EA	GOOG	MSFT	NVDA	SQ	WMT
Date
2021-04-12	59.889999	114.389999	139.680893	112.739502	250.827927	151.826202	265.200012	134.979538
2021-04-13	63.009998	120.879997	140.332626	113.363503	253.356720	156.523056	273.230011	134.564331

In [7]:

X.index = X.index.tz_localize(None)

In [8]:

df = pd.concat([y,X],axis=1).dropna()
df.head()

Out[8]:

	BTC	AI	CRSP	EA	GOOG	MSFT	NVDA	SQ	WMT
Date
2021-04-12	59893.453125	59.889999	114.389999	139.680893	112.739502	250.827927	151.826202	265.200012	134.979538
2021-04-13	63503.457031	63.009998	120.879997	140.332626	113.363503	253.356720	156.523056	273.230011	134.564331
2021-04-14	63109.695312	68.459999	123.110001	139.838898	112.741997	250.514282	152.505051	258.399994	134.516068
2021-04-15	63314.011719	66.500000	123.849998	140.984390	114.833000	254.346634	161.092590	263.079987	135.327118
2021-04-16	61572.789062	66.790001	118.559998	139.285889	114.888000	255.562057	158.849014	256.100006	135.761581

2.2 Analysis¶

Create training dataframe

Apply "Differencing" technique to make the data stationery

Visualize to check the correlation between "Treatment" and "Confounders"

Dropp least correlated confounders. Here the minimum threshold is 0.4

In [9]:

training_df = df[df.index <= training_end]
training_df.head()

Out[9]:

	BTC	AI	CRSP	EA	GOOG	MSFT	NVDA	SQ	WMT
Date
2021-04-12	59893.453125	59.889999	114.389999	139.680893	112.739502	250.827927	151.826202	265.200012	134.979538
2021-04-13	63503.457031	63.009998	120.879997	140.332626	113.363503	253.356720	156.523056	273.230011	134.564331
2021-04-14	63109.695312	68.459999	123.110001	139.838898	112.741997	250.514282	152.505051	258.399994	134.516068
2021-04-15	63314.011719	66.500000	123.849998	140.984390	114.833000	254.346634	161.092590	263.079987	135.327118
2021-04-16	61572.789062	66.790001	118.559998	139.285889	114.888000	255.562057	158.849014	256.100006	135.761581

In [10]:

differencing = training_df.pct_change().dropna()
differencing.head()

Out[10]:

	BTC	AI	CRSP	EA	GOOG	MSFT	NVDA	SQ	WMT
Date
2021-04-13	0.060274	0.052095	0.056736	0.004666	0.005535	0.010082	0.030936	0.030279	-0.003076
2021-04-14	-0.006201	0.086494	0.018448	-0.003518	-0.005482	-0.011219	-0.025670	-0.054277	-0.000359
2021-04-15	0.003237	-0.028630	0.006011	0.008192	0.018547	0.015298	0.056310	0.018111	0.006029
2021-04-16	-0.027501	0.004361	-0.042713	-0.012047	0.000479	0.004779	-0.013927	-0.026532	0.003210
2021-04-19	-0.094986	-0.070220	-0.029690	-0.010067	0.002019	-0.007671	-0.034611	-0.042054	-0.006400

💡 In 90% of the cases time series datas are non stationary. So we are assuming this dataframe containing non stationary data

In [11]:

sns.heatmap(data=differencing.corr(),
            annot=True,
            fmt=".1g",
            cmap="Blues")
plt.show()

No description has been provided for this image

In [12]:

final_df = df.drop(columns=["EA", "GOOG", "WMT"])
final_df.head(2)

Out[12]:

	BTC	AI	CRSP	MSFT	NVDA	SQ
Date
2021-04-12	59893.453125	59.889999	114.389999	250.827927	151.826202	265.200012
2021-04-13	63503.457031	63.009998	120.879997	253.356720	156.523056	273.230011

2.3 Model Creation¶

Define Pre & Post period

Create the impact model

Summarise the model

In [13]:

pre_period = [training_start,training_end]
post_period = [treatment_start, treatment_end]

In [14]:

impact = CausalImpact(data = final_df,
             pre_period = pre_period,
             post_period = post_period)
sns.set_theme()
impact.plot(figsize=(23,14))
plt.show()

C:\Users\Saiyan\miniconda3\envs\myenv\Lib\site-packages\statsmodels\tsa\base\tsa_model.py:473: ValueWarning: No frequency information was provided, so inferred frequency B will be used.
  self._init_dates(dates, freq)
C:\Users\Saiyan\miniconda3\envs\myenv\Lib\site-packages\statsmodels\base\optimizer.py:18: FutureWarning: Keyword arguments have been passed to the optimizer that have no effect. The list of allowed keyword arguments for method lbfgs is: m, pgtol, factr, maxfun, epsilon, approx_grad, bounds, loglike_and_score, iprint. The list of unsupported keyword arguments passed include: standardize, nseasons. After release 0.14, this will raise.
  warnings.warn(
C:\Users\Saiyan\miniconda3\envs\myenv\Lib\site-packages\statsmodels\tsa\statespace\representation.py:374: FutureWarning: Unknown keyword arguments: dict_keys(['alpha']).Passing unknown keyword arguments will raise a TypeError beginning in version 0.15.
  warnings.warn(msg, FutureWarning)
C:\Users\Saiyan\miniconda3\envs\myenv\Lib\site-packages\statsmodels\tsa\base\tsa_model.py:473: ValueWarning: No frequency information was provided, so inferred frequency B will be used.
  self._init_dates(dates, freq)

💡 Since we are working with stock datas, we will not consider the "Cumelitive Effect" and only conside the "Relative Effect"

In [15]:

print(impact.summary())

Posterior Inference {Causal Impact}
                          Average            Cumulative
Actual                    43971.47           263828.82
Prediction (s.d.)         49543.16 (885.73)  297258.94 (5314.37)
95% CI                    [47734.0, 51206.0] [286404.02, 307235.99]
Absolute effect (s.d.)    -5571.69 (885.73)  -33430.12 (5314.37)
95% CI                    [-7234.53, -3762.53][-43407.17, -22575.2]
Relative effect (s.d.)    -11.25% (1.79%)    -11.25% (1.79%)
95% CI                    [-14.6%, -7.59%]   [-14.6%, -7.59%]
Posterior tail-area probability p: 0.0
Posterior prob. of a causal effect: 100.0%
For more details run the command: print(impact.summary(‘report’))

2.4 Interpretation¶

The average value of the response variable during the post-intervention period was around 43971.47. If the intervention had not taken place, we would have expected an average response of 49543.17.The 95% interval for this expected value ranges from 47823.92 to 51362.75. That means if Elon Musk did not make that tweet the Bitcoin avarage price would be 49543.17 around that period, but because of the tweet the actual average Bitcoin price fall to 43971.47

By subtracting this expected value from the observed response, we estimate the causal effect of the intervention on the response variable to be -5571.7, with a 95% interval of -7391.28 to -3852.45. That indicates there is an negative effect on the Bitcoin price wich is around -5571.7

In relative terms, the response variable showed a decrease of approximately 11.25%. The 95% interval for this percentage change is from -14.92% to -7.78%. This indicates that the observed negative effect during the intervention period is statistically significant.

The probability of obtaining this effect by chance alone, as indicated by the Bayesian one-sided tail-area probability (p-value), is very small at 0.0. Therefore, the causal effect can be considered statistically significant.

3 Conclusion¶

By applying the Google Causal Impact we are able to prove there is a negative effect between the Elon Musk's Bitcoin tweet and the Bitcoin price fall. Understanding and analyzing causal relationships in cryptocurrency markets are crucial for investors, traders, policymakers, and researchers alike. By leveraging techniques like Google Causal Impact, we can gain a deeper understanding of the dynamics within these markets and make more informed decisions. This analysis serves as a reminder of the interconnectedness of social media, public figures, and the cryptocurrency landscape, and the potential implications of their interactions on market behavior.

Overall, the case study showcases the power of data analysis techniques like Google Causal Impact in studying and quantifying the impact of specific events or interventions on cryptocurrency prices. This knowledge can provide valuable insights for market participants and contribute to a more comprehensive understanding of the complex and dynamic nature of the cryptocurrency market.

Elon Musk vs Bitcoin

A "Google Causal Impact" Analysis Between Elon Musk Tweet's & BTC Downfall