Time Series Forecasting
business analytics
slug: topic-map-business-analytics-time-series-forecasting
Vocabulary:
- time_series: Sequence of data points ordered by time
- trend: Long-term upward or downward movement
- seasonality: Regular, predictable patterns that repeat over time
- cycle: Longer-term fluctuations without fixed periodicity
- noise: Random, irregular variations
- stationarity: Statistical properties (mean, variance) constant over time
- autocorrelation: Correlation of series with lagged version of itself
- lag: Time offset (e.g., lag-1 = previous period, lag-12 = same month last year)
- differencing: Subtracting previous value to remove trend
- decomposition: Separating time series into trend, seasonal, and residual components
- smoothing: Techniques to reduce noise (moving average, exponential smoothing)
- periodicity: Length of seasonal cycle (e.g., 12 for monthly data with yearly season)
Concepts:
- additive_model: Y = Trend + Seasonal + Error (seasonal amplitude constant)
- multiplicative_model: Y = Trend × Seasonal × Error (seasonal amplitude grows with trend)
- classical_decomposition: Breaking into trend, seasonal, irregular components
- trend_estimation: Using moving averages or regression to estimate trend
- seasonal_indices: Average seasonal pattern across all cycles
- deseasonalization: Removing seasonal pattern to see underlying trend
- forecast_horizon: How many periods ahead to predict
- holdout_validation: Reserving recent periods to test forecast accuracy
- ensemble_forecasting: Combining multiple methods for robustness
Concepts_advanced:
- stl_decomposition: Seasonal-Trend decomposition using LOESS
- fourier_decomposition: Using sine/cosine waves for seasonal patterns
- changepoint_detection: Identifying structural breaks in time series
- intervention_analysis: Accounting for known events (promotions, holidays)
- hierarchical_forecasting: Ensuring forecasts at different levels sum correctly
- reconciliation: Adjusting forecasts to maintain cube structure integrity
Procedures:
- validate_time_series: Ensure regular time intervals, no missing periods
- calculate_moving_average: Smooth data with rolling window
- estimate_trend: Linear regression or polynomial fit
- calculate_seasonal_indices: Average each season across years
- deseasonalize_data: Divide by (multiplicative) or subtract (additive) seasonal indices
- calculate_residuals: Observed - (Trend + Seasonal)
- test_stationarity: Augmented Dickey-Fuller test or similar
- calculate_autocorrelation: Correlation at different lags
- identify_periodicity: Find dominant cycle length (ACF peaks)
- prepare_forecast_features: Create lag variables, seasonal dummies, trend variable
Procedures_detailed:
- centered_moving_average: For even periods, average adjacent pairs
- seasonal_index_calculation:
- Calculate moving average
- Detrend data (observed / MA for multiplicative)
- Average each seasonal position across cycles
- Normalize indices to sum/average to target
- additive_decomposition: Y = MA + (Y - MA) seasonally averaged + residual
- multiplicative_decomposition: Y = MA × (Y / MA) seasonally averaged × residual
- exponential_smoothing: forecast = α × actual + (1-α) × previous_forecast
Topics:
- demand_forecasting
- revenue_projection
- capacity_planning
- inventory_optimization
- budget_planning
- workforce_scheduling
- trend_analysis
- seasonal_pattern_discovery
- anomaly_contextualization
- what_if_scenario_planning
Categories:
- temporal_analysis
- predictive_analytics
- pattern_decomposition
- forecast_preparation
- trend_identification
Themes:
- understanding_patterns: Separate signal from noise
- future_visibility: Enable forecasting by understanding structure
- seasonal_awareness: Account for predictable cycles
- trend_identification: Know if heading up, down, or stable
Trends:
- automated_model_selection: AI picks best forecasting method
- prophet_style_forecasting: Facebook Prophet for business time series
- deep_learning_forecasting: LSTMs, transformers for complex patterns
- probabilistic_forecasting: Prediction intervals not just point forecasts
- causal_forecasting: Incorporating external drivers (weather, events, promotions)
Use_cases:
- retail: "Decompose monthly sales into trend (+2%/month), seasonality (Dec=180 index), residual"
- saas: "MRR shows linear trend +$50K/month, plus 12-month renewal cycle pattern"
- manufacturing: "Production volume has 7% annual growth trend plus quarterly seasonality (Q4 index=125)"
- hospitality: "Occupancy has weekly seasonality (weekend index=140) plus yearly summer peak"
- ecommerce: "Traffic shows day-of-week pattern (Sun index=75) plus holiday spikes"
- finance: "Loan applications trend upward 3%/year with spring peak (Apr index=115)"
- healthcare: "ER visits show weekly pattern (Mon index=110) plus flu season in Q1"
- utilities: "Energy usage has strong annual seasonality (summer/winter peaks) plus gradual growth trend"