CryptoVision : Advanced and Scalable Crypto Price Forecasting

Scalabiltiy Results

Percentage Improvement in Speed: 7784.80%

PySpark Output was 78.85 times more scalable than Pandas.

Model Comarison Results

Here I have tried 20 models for time series forecasting on Bitcoin dataset from 17-7-2010 to 9-9-2024 and Best performance was given be Ensemble Model

Scalability was done by predicting on Top 50 Crypto creating a PandasUDF integrating it with Pyspark

Statistical Models

1. AR
2. MA
3. ARMA
4. AIRMA
5. SARIMA
6. Naive Forecast
7. AutoARIMA
8. ExponentialSmoothing (Requires Smoothening)

Machine Learning Methods

9. Random Forest (TF-DF)
10. Gradient Boosted Trees (TF-DF)
11. Prophet (Facebook Kats) (Requires Smoothening)
12. Dense Model (Window = 7, Horizon = 1)
13. Dense Model (Window = 30, Horizon = 1)
14. Dense Model (Window = 30, Horizon = 7)
15. Conv1D (Window = 7, Horizon = 1)
16. LSTM (Window = 7, Horizon = 1)
17. Dense (Multivariate Time series)
18. N-BEATs Algorithm
19. Ensemble
20. Simple ANN Model (Future Predictions)

Preprocessing Steps

1. Checking if series is stationary Using Augmented Dickey Fuller Test
2. Making the series stationary Using Differencing
3. Plotting ACF and PACF plots According to the plots mostly both the plots were same and the first lag was negative second lag onwards many lags were below the threshold

Conclusion

1. The test dataset was of 362 days (around 1 year) this is the main reason why the forecast of FB-Prophet, ARIMA, ARMA, AR, MA etc are not so good.