Sales Forecasting with Time Series Analysis

Technical Highlights

Time Series Analysis: Comprehensive comparison ARIMA (time series specialist) vs Linear Regression (simple baseline) dengan systematic evaluation.

Model Optimization: GridSearchCV untuk hyperparameter tuning, Ridge regression dengan L2 regularization untuk better generalization.

Surprising Results: Simple Linear Regression outperformed ARIMA by 41% (MAE $3,726 vs $5,800), demonstrating importance of model selection based on data characteristics.

Interactive Dashboard: Streamlit application dengan forecasting interface, confidence intervals, dan business insights visualization.

Skills Demonstrated

Time Series Analysis

• ▸ARIMA Modeling: AutoRegressive Integrated Moving Average
• ▸Parameter Selection: AIC-based optimization untuk (p,d,q) parameters
• ▸Stationarity Testing: Augmented Dickey-Fuller test
• ▸Seasonal Decomposition: Trend, seasonality, residual analysis

Machine Learning

• ▸Linear Models: Linear Regression, Ridge Regression
• ▸Hyperparameter Tuning: GridSearchCV dengan cross-validation
• ▸Model Evaluation: MAE, RMSE, R-squared
• ▸Regularization: L2 regularization untuk prevent overfitting

Data Science

• ▸Exploratory Data Analysis: Distribution analysis, pattern detection
• ▸Feature Engineering: Time-based features, aggregation strategies
• ▸Data Resampling: Daily → Weekly aggregation untuk noise reduction
• ▸Validation Strategy: Time-series split (90-10)

Software Engineering

• ▸Python Development: Clean, modular code dengan proper documentation
• ▸Interactive Dashboard: Streamlit untuk user-friendly forecasting interface
• ▸Visualization: Matplotlib, Seaborn untuk insights communication
• ▸Deployment: Streamlit Cloud dengan automated deployment

Model Comparison Results

| Model | MAE | RMSE | vs ARIMA | |-------|-----|------|----------| | ARIMA (5,1,0) | $5,800 | $8,178 | Baseline | | Linear Regression | $3,726 | $4,792 | 41% better | | Ridge (α=0.1) | $3,676 | $5,049 | Best |

Winner: Ridge Regression dengan alpha=0.1

Why Linear Regression Won?

Data Characteristics:

• ▸No strong seasonality (ARIMA's strength wasted)
• ▸Short time series (13 weeks, not enough untuk ARIMA)
• ▸Clear linear trend (perfect untuk LR)
• ▸Simple model prevents overfitting

Lesson: Model complexity ≠ Better performance. Always test simple baselines.

Technical Implementation

ARIMA Parameter Selection:

# Test multiple (p,d,q) combinations
for p in range(6):
    for d in range(2):
        for q in range(3):
            model = ARIMA(train, order=(p,d,q))
            aic = model.fit().aic
            # Select best based on AIC

Ridge Regression Tuning:

param_grid = {'alpha': [0.1, 1.0, 10.0, 100.0]}
grid_search = GridSearchCV(Ridge(), param_grid, cv=5)
best_model = grid_search.fit(X_train, y_train)

Feature Engineering

Temporal Features:

• ▸Month, Day, Weekday extraction
• ▸Hour-based patterns (peak hours: 13:00-15:00, 19:00-20:00)

Aggregation Strategy:

• ▸Daily sales → Weekly sales (noise reduction)
• ▸Smoothing untuk better trend detection

Derived Metrics:

• ▸Rolling averages
• ▸Growth rates
• ▸Seasonal indicators

Business Insights

Forecast Results: 8% growth over next 12 weeks

Actionable Recommendations:

• ▸Inventory Planning: Increase stock by 10% (with buffer)
• ▸Staffing Optimization: More staff during peak hours (13:00-15:00, 19:00-20:00)
• ▸Promotion Planning: Target slow days untuk boost sales

Expected Impact:

• ▸30% reduction in stockouts
• ▸20% improvement in labor efficiency
• ▸15% increase in slow-day sales

Technical Challenges Solved

Short Time Series: Only 13 weeks data. ARIMA needs 50+ observations. Solution: Use simple models yang nggak overfit.

High Variability: Daily sales std $1,842. Solution: Weekly aggregation untuk smooth noise.

Model Selection: Systematic comparison dengan consistent metrics. Don't assume ARIMA is always best.

Hyperparameter Tuning: GridSearchCV dengan cross-validation untuk find optimal Ridge alpha.

Deployment: Interactive Streamlit dashboard dengan:

• ▸Historical data visualization
• ▸Forecast horizon selection (1-52 weeks)
• ▸Confidence intervals
• ▸Business metrics (total forecast, avg weekly, growth rate)

Live Demo: https://sales-forecasting-fauza.streamlit.app/

Read Full Story: Blog Post