Blog

Time Series Forecasting

22 January 2025

Time series forecasting is a technique for predicting future values based on past data collected over time. This is useful in economics, sales, weather prediction, stock prices, and many more fields. In this article, we introduce three popular forecasting models: ARIMA, Facebook Prophet, and LSTM neural networks.

What is a Time Series?

A time series is a sequence of data points measured at regular time intervals (e.g., daily sales, hourly temperatures, monthly stock prices). The key idea is that each value often depends on its previous values.

Challenges in Time Series Forecasting

Trends: Long-term increase or decrease in the data.
Seasonality: Regular, repeated patterns (like more sales on weekends).
Noise: Random changes or outliers that make forecasting harder.
Missing values: Sometimes we don’t have data for every time point.

Popular Time Series Forecasting Models

ARIMA (AutoRegressive Integrated Moving Average)

Traditional statistical model for time series forecasting.
Works well for many types of real-world data.
Captures trends and patterns using the past values and errors.
Best if data shows consistent trends or seasonal behavior.

Example (using Python's statsmodels):

from statsmodels.tsa.arima.model import ARIMA
import pandas as pd

# Example: Forecast a monthly time series
data = pd.read_csv('monthly-sales.csv')   # Must have 'Month' and 'Sales' columns
model = ARIMA(data['Sales'], order=(2,1,2))
model_fit = model.fit()
forecast = model_fit.forecast(steps=12)
print(forecast)

Prophet (by Facebook)

Flexible and easy-to-use forecasting tool from Facebook.
Handles seasonality, holidays, and missing values well.
Works great for business or web data with clear seasonal trends.

Example (with Python's prophet package):

from prophet import Prophet
import pandas as pd

# Load data: must have columns 'ds' (date) and 'y' (value)
df = pd.read_csv('monthly-sales.csv')
df = df.rename(columns={'Month': 'ds', 'Sales': 'y'})

model = Prophet()
model.fit(df)
future = model.make_future_dataframe(periods=12, freq='M')
forecast = model.predict(future)
print(forecast[['ds', 'yhat', 'yhat_lower', 'yhat_upper']].tail(12))

LSTM (Long Short-Term Memory Networks)

A type of recurrent neural network (RNN) good at learning from sequential data.
Great for complex, non-linear time series (e.g., stock prices, IoT sensor data).
Usually needs more data and computing power to train compared to ARIMA/Prophet.
Works without strict assumptions about your data.

Simple LSTM Example (using TensorFlow and Keras):

from keras.models import Sequential
from keras.layers import LSTM, Dense
import numpy as np

# Simulated data: X_train shape (samples, time_steps, features)
X_train = np.random.rand(100, 12, 1)  # 100 samples, 12 time steps, 1 feature
y_train = np.random.rand(100)

model = Sequential()
model.add(LSTM(50, activation='relu', input_shape=(12, 1)))
model.add(Dense(1))
model.compile(optimizer='adam', loss='mse')
model.fit(X_train, y_train, epochs=30, batch_size=16)

This uses randomly generated data for illustration. For real use, reshape your own time series into sequences and targets.

Which Model Should You Choose?

ARIMA: Simple and powerful for many business and economic forecasts. Good for short-term trends and stationary data.
Prophet: Best if you have seasonal data and want quick, interpretable results (especially if you like visual plots).
LSTM: Choose for complex patterns in large datasets—especially if classical models do not work well.

Applications

Sales and demand forecasting
Energy consumption prediction
Stock market price prediction
Weather and climate modeling
Traffic and transport forecasting

Conclusion

Time series forecasting is a crucial tool for making decisions based on past data. With models like ARIMA, Prophet, and LSTM, you can analyze trends, forecast the future, and discover hidden patterns in your temporal data. Start with ARIMA or Prophet for easier projects, and explore LSTM for more complex tasks!