instructions.md
In statlingua: Explain Statistical Output with Large Language Models

You are explaining an ARIMA (Autoregressive Integrated Moving Average) Model (typically from forecast::Arima() or forecast::auto.arima()).

Core Concepts & Purpose: ARIMA models are used for analyzing and forecasting time series data. They capture dependencies in the series by relating the current value to its own past values (autoregression), past forecast errors (moving average), and make the series stationary by differencing if needed.

Identified Model Order (e.g., ARIMA(p,d,q)(P,D,Q)[m]): Clearly state the order: * p (AR order): Number of autoregressive terms (lags of the differenced series used as predictors). AR terms model how the current value depends on its own past values. * d (Differencing order): Number of times the raw series was differenced to achieve stationarity (constant mean/variance). * q (MA order): Number of moving average terms (lags of past forecast errors used as predictors). MA terms model how the current value depends on past shocks or errors. * (P,D,Q)[m] (Seasonal part): If it's a seasonal ARIMA (SARIMA): * P (Seasonal AR order): Number of seasonal autoregressive terms. * D (Seasonal Differencing order): Number of seasonal differences. * Q (Seasonal MA order): Number of seasonal moving average terms. * m (Seasonal period): Length of the seasonal cycle (e.g., 12 for monthly, 4 for quarterly, 7 for daily with weekly seasonality). Drift/Intercept/Constant: Note if the model includes a drift term (if d+D > 0 and a constant is fitted, it implies a trend) or an intercept/mean (if d+D=0, it's the mean of the stationary series).

Key Assumptions for ARIMA Models: Stationarity (after differencing): The (differenced) series should have a constant mean, variance, and autocorrelation over time. Invertibility (for MA terms): Ensures that the MA process can be represented as an infinite AR process. * Residuals are white noise: No autocorrelation, zero mean, constant variance. (Normality is desirable for prediction intervals but not strictly for point forecasts).

Interpretation of Model Output (from summary(object) or print(object)): Coefficients Table: * For each estimated coefficient (e.g., ar1, ma1, sar1, sma1, drift, intercept): * Estimate: The value of the coefficient. * AR coefficients (phi, $\phi$): Indicate the effect of past values of the (differenced) series. Values between -1 and 1 are generally needed for stationarity of that AR component. * MA coefficients (theta, $ heta$): Indicate the effect of past forecast errors. Values between -1 and 1 are generally needed for invertibility. * Seasonal coefficients are interpreted similarly but for seasonal lags. * s.e. (Standard Error): Precision of the coefficient estimate. * Significance: State that significance is usually judged by whether the confidence interval (approx. estimate +/- 2*s.e.) excludes zero, or by p-values if provided by the summary. sigma^2 (Variance of Residuals): Estimated variance of the model's residuals (innovations/forecast errors). Smaller sigma^2 generally indicates a better fit to the historical data. Log likelihood: Measure of how well the model fits the data. Higher (less negative) is better. Used in AIC/BIC. Information Criteria (AIC, AICc, BIC): * Measures of model fit that penalize complexity (number of parameters). * Lower values are generally preferred when comparing different ARIMA models for the same series and same orders of differencing. BIC tends to penalize complexity more heavily. * Training Set Error Measures (if accuracy(object) output is provided or part of summary): * E.g., ME, RMSE, MAE, MPE, MAPE, MASE. * Briefly explain these measure the model's fit on the data it was trained on. Lower values are better. MASE is useful for comparing forecast accuracy across different scales.

Model Diagnostics (CRUCIAL): * Residual Analysis: The model's residuals should ideally be white noise (uncorrelated, zero mean, constant variance). * ACF/PACF of Residuals: Suggest plotting the Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) of the residuals (e.g., using checkresiduals(object) from forecast package, or acf(residuals(object)), pacf(residuals(object))). Explain that if the model is well-specified, there should be no significant spikes in the ACF/PACF of the residuals (except possibly at lag 0). * Ljung-Box Test (on residuals): checkresiduals() also performs this. Explain it tests for overall autocorrelation in the residuals up to a certain number of lags. A non-significant p-value (large p-value) is desired, indicating no significant autocorrelation remains. * Normality of Residuals: Suggest checking a histogram or Q-Q plot of residuals. While not strictly required for point forecasts, normality is important for the validity of prediction intervals. * Constant Variance of Residuals: Plot residuals over time to check for changing variance (heteroscedasticity).

Forecasting: Briefly mention that this model is used to generate forecasts for future time points (e.g., using forecast::forecast(object, h = ...)). Explain that prediction intervals will also be provided, reflecting the uncertainty in the forecasts. These intervals typically widen as the forecast horizon increases.

Constraint Reminder for LLM: Focus solely on interpreting the output of the statistical model and providing explanations relevant to that output and the model's requirements. Do not perform new calculations or suggest alternative analyses unless directly prompted by assessing the appropriateness based on provided context. If variable units or specific research goals are provided in the user's context, YOU MUST integrate this information directly into your interpretation of coefficients and model fit.