This repository hosts a collection of notebooks featuring an in-depth analysis of ten cryptocurrencies, along with a comparative study of predictions made by various models specifically for BTC. The following models were evaluated:
- ARIMA
- XGBoost
- LSTM GRU
- Prophet
Feel free to ⭐ and clone this repo 😉
In the first stage, a comparison of 10 cryptocurrencies was carried out. For that, data from yfinance from January 2019 to June 2023 were downloaded and the following features were extracted:
- Date: Trade date
- Open: Opening stock price
- High: Peak price during the trading day
- Low: Lowest price during the trading day
- Close: Closing stock price
- Adj Close: Adjusted closing price, accounting for splits, offerings, and dividends for accurate historical comparison
- Volume: Total trade volume for the day
The utilization of Adj Close allows for meaningful comparisons between cryptocurrencies, and these values are thoughtfully visualized through multiple subplots.
In the pursuit of enhanced insights, returns were computed, providing a measure of profitability over a specific period. Furthermore, cumulative returns were calculated to express the overall price change of a cryptocurrency over time.
Notably, a pattern emerges revealing varying degrees of correlation between different cryptocurrencies. BCH, ETH, EOS, and LTC exhibit a strong positive correlation, implying that when one of these cryptocurrencies experiences an increase, the others tend to follow suit. Conversely, BSV displays a weaker correlation with BNB and XRP.
The application of a moving average, also known as a rolling average, helped in smoothing out short-term price fluctuations to identify longer-term trends or cycles. For instance, a 7-day moving average captures short-term trends, while a 100-day rolling average highlights more substantial trends. The use of arithmetic mean was adopted in this context.
From the plotted chart, it is discernible that in May of 2021, the price dipped below the 50-day moving average, signifying a downward trend. Conversely, in August 2021, the price surged above the moving average, indicating an upward trend.
Further observations reveal that around May of 2021, the 20-day moving average crossed below the 100-day moving average, signaling a shift towards a downward trend and serving as a sell signal. Conversely, by August 2021, the 20-day moving average crossed above the 100-day moving average, marking an upward trend and presenting a buy signal.
The subsequent phase of analysis delved into an in-depth exploratory data analysis (EDA) of BTC, aiming to uncover valuable insights. This analysis period encompassed data obtained from Binance, spanning from January 2020 to May 2023, as Binance provides a more comprehensive dataset. The dataset from Binance included various data attributes such as:
- open_time: Opening time of the specified period
- open: Opening price of the trading instrument for the given period
- high: Highest price attained by the trading instrument during the specified period
- low: Lowest price reached by the trading instrument during the specified period
- close: Closing price of the trading instrument for the given period
- volume: Total trading volume, representing the quantity of the trading instrument, during the specified period
- close_time: Closing time of the specific period
- quote_volume: Total volume in terms of the quote asset, which determines the trading instrument's value
- count: Number of trades occurring during the specified period
- taker_buy_volume: Volume of the quote asset bought by market takers during the period
- taker_buy_quote_volume: Volume of the base asset bought by market takers during the period
To ensure a comprehensive and accurate dataset, the following actions were taken:
- Identification and handling of missing data points and their corresponding values
- Utilization of interpolation to fill in the gaps with the mean values
Upon obtaining the complete dataset (saved under BTCUSDT_final_clean.csv), the primary focus shifted toward the analysis of the closing prices. It was noted that in both 2020 and 2022, the cryptocurrency market experienced two significant crashes attributed to the impact of the COVID pandemic and the collapse of the FTX exchange.
Another insightful indicator employed was the lag plot. The visuals illustrated that lags of 1 and 2 days exhibited minimal noise, suggesting the potential for time series resampling. Subsequently, a 1-day resampling interval was chosen for further exploration.
Among the models examined later in the analysis, ARIMA was a key focus. This model necessitated several steps including time series decomposition, autocorrelation assessment (ACF and PACF), and stationarity checks (KPSS and ADF). The findings revealed the absence of seasonality in the series, with both p and q values being 1 each, and a differencing level of 1 was employed to attain stationarity. Intriguingly, the p, q, and d values determined by the auto_arima method were (1,1,0).
Thorough evaluations were conducted on the following models:
- ARIMA
- XGBoost
- LSTM GRU
- Prophet
Overall, all models demonstrated impressive performance, yielding highly favorable outcomes. The actual versus predicted price curve displayed notable similarity across all models, as illustrated in the example below, showcasing the LSTM GRU model's projection.
A concise summary of the achieved results is presented in the table below, showcasing key metrics for each model. It is noteworthy that while Prophet achieved the highest R-squared (R2) value, indicating strong goodness-of-fit, ARIMA exhibited the best Mean Absolute Error (MAE) and Mean Squared Error (MSE) values, highlighting its accuracy in prediction.
| Metric | ARIMA | LSTM | Prophet | XGBoost |
|---|---|---|---|---|
| MAE | 77 | 153 | 125 | 709 |
| MSE | 131 | 250 | 157 | 1174 |
| R2 | 98.86 | 99.69 | 99.88 | 93.08 |
These findings underscore the effectiveness of the models in capturing BTC price trends, with each model showcasing unique strengths in specific metrics. This comprehensive evaluation contributes to a robust understanding of the models' capabilities and assists in informed decision-making.