Forecast Sales Accurately with Advanced Models

This is a sales-forecasting skill covering statistical and machine-learning models, accuracy measurement, and scenario planning, built on time-series foundations (trend, seasonality, cyclical, and irregular components) and a data-quality bar of at least 24 months of clean historical data. It implements four model families: moving averages (simple and weighted), exponential smoothing (via statsmodels' ExponentialSmoothing, with trend/seasonal/seasonal_periods configuration), ARIMA (with stationarity testing via adfuller and a grid search over p/q parameters to minimize AIC), and a machine-learning approach using RandomForestRegressor with engineered lag, rolling-statistics, calendar, and seasonal-decomposition features. Forecast accuracy is measured with a dedicated function covering four metrics at once:

def calculate_forecast_accuracy(actual, predicted):
    """Calculate comprehensive forecast accuracy metrics"""
    mae = mean_absolute_error(actual, predicted)
    mse = mean_squared_error(actual, predicted)
    rmse = np.sqrt(mse)
    
    # Mean Absolute Percentage Error
    mape = np.mean(np.abs((actual - predicted) / actual)) * 100
    
    # Mean Absolute Scaled Error
    naive_forecast = actual[:-1]  # Naive forecast is previous period's actual
    naive_mae = mean_absolute_error(actual[1:], naive_forecast)
    mase = mae / naive_mae if naive_mae != 0 else np.inf
    
    return {
        'MAE': mae,
        'RMSE': rmse,
        'MAPE': mape,
        'MASE': mase
    }

Use this skill when building a forecasting pipeline that needs to move from simple baselines to more sophisticated models methodically - the model-selection strategy explicitly says to start simple with moving averages and exponential smoothing, validate with walk-forward validation, and only add ensembles or ARIMA/ML approaches once the simpler models are understood. It also covers scenario planning (applying named growth-rate and seasonal-boost adjustments to a base forecast to produce optimistic, pessimistic, and realistic variants) and forecast-bias analysis (whether a model systematically over- or under-forecasts). It is not a one-shot prediction tool - it assumes at least 24 months of historical data and ongoing model recalibration, so it isn't suited to a business without enough sales history to establish a real pattern.

Input is historical sales data (with a minimum 24-month history, segmented by product line, region, or customer type for granular accuracy) plus, for the ML approach, engineered features like lagged sales values, rolling means and standard deviations, and calendar fields. Output is a forecast for a specified number of future periods, accuracy metrics (MAE, RMSE, MAPE, and MASE against a naive baseline), a bias-direction read (over- versus under-forecasting), and - via the scenario-planning class - multiple named forecast variants such as optimistic, pessimistic, and realistic, with per-month seasonal adjustments like a holiday boost.

Built on pandas, numpy, scikit-learn (RandomForestRegressor, StandardScaler, accuracy metrics), and statsmodels (ExponentialSmoothing, ARIMA, adfuller, seasonal_decompose); designed to connect to CRM systems for pipeline data and to feed a forecast-performance dashboard tracking multiple models' accuracy over time.

Sales operations and data teams building or maintaining a sales-forecasting pipeline who want a methodical path from simple statistical baselines to ARIMA or ML models, with built-in accuracy measurement and scenario planning, rather than a single fixed forecasting method.