Skip to content

Time_Series_Forecasting.md

Latest commit

 

History

History
94 lines (73 loc) · 3.16 KB

Time_Series_Forecasting.md

File metadata and controls

94 lines (73 loc) · 3.16 KB
Raw

Skill: Time Series Forecasting

Purpose

To predict future values based on previously observed time-ordered data, considering trends, seasonality, and exogenous variables.

When to Use

  • When forecasting sales, inventory, stock prices, or resource demand
  • When identifying anomalies in system metrics over time
  • When working with IoT sensor data

Procedure

1. Identify the Time Series Characteristics

Analyze the data for:

  • Trend: Long-term upward or downward movement.
  • Seasonality: Repeating patterns at fixed intervals (daily, weekly, yearly).
  • Stationarity: Whether statistical properties (mean, variance) change over time.

2. Choose the Forecasting Model

Select an appropriate algorithm based on data complexity:

  • ARIMA / SARIMA: Traditional statistical models. Best for univariate data with clear seasonality.
  • Prophet (by Meta): Excellent for business time series with daily observations and strong seasonal effects.
  • XGBoost / LightGBM: Effective for tabular time series with many exogenous features.
  • LSTMs / Transformers (e.g., Temporal Fusion Transformer): Best for complex, non-linear relationships and long sequences.

3. Prophet Implementation Example

Setup and Fitting:

from prophet import Prophet
import pandas as pd

# Data must have 'ds' (datestamp) and 'y' (target) columns
df = pd.read_csv('sales_data.csv')
df.rename(columns={'date': 'ds', 'sales': 'y'}, inplace=True)

# Initialize Prophet model
m = Prophet(
    yearly_seasonality=True,
    weekly_seasonality=True,
    daily_seasonality=False
)

# Add custom holidays or exogenous variables if necessary
m.add_country_holidays(country_name='US')

# Fit the model
m.fit(df)

Forecasting:

# Create future dates
future = m.make_future_dataframe(periods=365)

# Predict future values
forecast = m.predict(future)

# Plotting results
fig1 = m.plot(forecast)
fig2 = m.plot_components(forecast)

4. Evaluation Metrics

Use appropriate error metrics for time series:

  • MAE (Mean Absolute Error): Easy to interpret, robust to outliers.
  • RMSE (Root Mean Squared Error): Penalizes large errors heavily.
  • MAPE (Mean Absolute Percentage Error): Useful for comparing relative performance across different scales.
from sklearn.metrics import mean_absolute_error, mean_squared_error
import numpy as np

mae = mean_absolute_error(y_true, y_pred)
rmse = np.sqrt(mean_squared_error(y_true, y_pred))

5. Cross-Validation

Do not use standard K-Fold cross-validation. Use Time Series Split to respect temporal order.

from sklearn.model_selection import TimeSeriesSplit

tscv = TimeSeriesSplit(n_splits=5)
for train_index, test_index in tscv.split(X):
    X_train, X_test = X[train_index], X[test_index]
    y_train, y_test = y[train_index], y[test_index]
    # Train and evaluate model

Best Practices

  • Always plot your data before modeling. Visual inspection reveals obvious anomalies or structural breaks.
  • Use baseline models (like naive forecasting or moving average) to establish a performance floor before moving to complex models.
  • Handle missing values carefully; avoid interpolating over large gaps without justification.