Machine Learning: ARIMA + SARIMA

Description

The ARIMA (Autoregressive Integrated Moving Average) and SARIMA (Seasonal ARIMA) are advanced statistical models that use machine learning to forecast future price movements. It uses autoregression to find the relationship between observed data and its lagged observations. The data is differenced to make it more predictable. The MA component creates a dependency between observations and residual errors. The parameters are automatically adjusted to market conditions.

Differences

ARIMA - This excels at identifying trends in the form of directions
SARIMA - Incorporates seasonality. It's better at capturing patterns previously seen

How To Use

1. Model: Determine if you want to use ARIMA (better for direction) or SARIMA (better for overall prediction). You can click on the 'Show Historic Prediction' to see the direction of the previous candles. Green = forecast ending up, red = forecast ending down
2. Metrics: The RMSE% and MAPE are 10 day moving averages of the first 10 predictions made at candle close. They're error metrics that compare the observed data with the predicted data. It is better to use them when they're below 8%. Higher timeframes will be higher, as these models are partly mean-reverting and higher TFs tend to trend more. Better to compare RMSE% and MAPE with similar timeframes. They naturally lag as data is being collected
3. Parameter selection: The simpler, the better. Both are used for ARIMA(1,1,1) and SARIMA(1,1,1)(1,1,1)5. Increasing may cause overfitting
4. Training period: Keep at 50. Because of limitations in pine, higher values do not make for more powerful forecasts. They will only criminally lag. So best to keep between 20 and 80