Machine Learning: ARIMA + SARIMA

The Machine Learning: ARIMA + SARIMA indicator provides advanced statistical forecasting by utilizing autoregressive models to project future price movements. It automatically differences market data to achieve stationarity and adjusts parameters based on current market conditions to provide predictive insights into potential trends and seasonal patterns.

Usage

The Usage section describes how the script can be used to interpret market behavior and forecast future outcomes.

• Model Selection: Users can choose between two primary models:
  • ARIMA: Optimized for identifying general trends and directions.
  • SARIMA: Incorporates seasonality, making it more effective at capturing recurring historical patterns.
• Directional Visualization: By enabling "Show Historic Prediction," the script displays the direction of previous forecasts. Green indicates a predicted upward movement, while red indicates a predicted downward movement.
• Accuracy Metrics: The indicator tracks performance via RMSE% and MAPE (10-day moving averages). These error metrics compare observed data with predicted data. Values below 8% generally suggest higher reliability. Note that higher timeframes often show higher error rates due to trending behavior vs. the mean-reverting nature of these models.
• Training Period: The training period is best kept between 20 and 80 (defaulting to 50). Using excessively high values may introduce significant lag and reduce the responsiveness of the machine learning model.

Details

The script implements the Autoregressive Integrated Moving Average (ARIMA) and its Seasonal counterpart (SARIMA) using Pine Script.

• Autoregression (AR): Finds relationships between observed data and lagged observations to recognize patterns.
• Integrated (I): Uses differencing to remove non-stationary elements (like trends) from the data, making it more predictable for the model.
• Moving Average (MA): Creates a dependency between an observation and a residual error from a moving average model applied to lagged observations.
• Seasonality (SARIMA only): Adds parameters to model recurring cycles (Seasonal Period) to account for time-dependent patterns often found in financial markets.

Settings

The settings are categorized into general model parameters and specific ARIMA/SARIMA configurations:

• Model Type: Selects between ARIMA (directional focus) or SARIMA (pattern focus).
• Forecast Length: Sets the number of bars into the future the model will predict (max 50).
• Training Period: Determines how many historical data points are used to train the model.
• Memory: Adjusts the weight given to older data (Low, High, or Auto).
• AR Order (p): Sets the Autoregressive order; higher values improve recognition but risk overfitting.
• Differencing (d): Sets the order of differencing to make data stationary.
• MA Order (q): Sets the Moving Average order to account for recent error sensitivity.
• Seasonal Period (s): Defines the number of bars in a seasonal cycle.
• Seasonal AR Order (P): Adjusts the seasonal autoregressive complexity.
• Seasonal Differencing (D): Applies differencing to seasonal cycles.
• Seasonal MA Order (Q): Models error terms specifically across different seasons.
• Show Historic Prediction: Toggles the visibility of previous directional forecasts on the chart.

FAQ

How do I interpret the RMSE% and MAPE values? These are error metrics that indicate the accuracy of the forecast. Lower percentages mean the model's predictions are closer to the actual price. It is generally recommended to look for values below 8% for higher confidence.

Why does the forecast look different on higher timeframes? Higher timeframes tend to trend more aggressively, while ARIMA/SARIMA models are partially mean-reverting. This can lead to larger deviations between predictions and actual price action on daily or weekly charts compared to intraday charts.

How do I get access to Machine Learning: ARIMA + SARIMA? You can get access on the LuxAlgo Library for charting platforms like TradingView, MetaTrader (MT4/MT5), and NinjaTrader for free.