Building a Complete Forecasting Pipeline

Until now, we learned individual concepts — trend, seasonality, models, and evaluation.

In real projects, these steps are never used in isolation. They are connected together into a forecasting pipeline.

This lesson shows how everything fits together — logically and visually.

A Real-World Scenario

Imagine you work for a food delivery company.

You are asked to forecast daily order volume so the company can:

• Schedule delivery staff
• Prepare inventory
• Avoid delays during peak days

Let’s build a simple but realistic forecasting pipeline for this situation.

Step 1: Understanding the Raw Data

Every pipeline starts with raw data.

Below is a simulated daily order count:

• Orders slowly increase as the business grows
• Weekly seasonality (weekends are busier)
• Some random noise

import numpy as np
import matplotlib.pyplot as plt

np.random.seed(4)
days = np.arange(140)
trend = 0.2 * days
seasonal = 10 * np.sin(2 * np.pi * days / 7)
noise = np.random.normal(0, 4, size=140)

orders = 80 + trend + seasonal + noise

plt.figure(figsize=(9,4))
plt.plot(orders)
plt.title("Daily Orders")
plt.show()

From this plot, we can already observe:

• A slow upward trend
• Clear weekly repetition
• Some randomness

This visual understanding is critical before modeling.

Step 2: Train–Test Split (Time-Aware)

We never train on future data.

So we split the series:

• Past → training
• Future → testing

train = orders[:120]
test = orders[120:]

This mirrors how forecasting works in the real world:

You only know the past when predicting the future.

Step 3: Choose a Baseline Model

Before using complex models, we start simple.

A common baseline is the naive forecast:

Tomorrow’s value ≈ today’s value.

forecast = np.repeat(train[-1], len(test))

This forecast:

• Ignores trend
• Ignores seasonality
• Sets a performance baseline

Any serious model must beat this.

Step 4: Evaluate the Forecast

We compare predictions against actual future values.

Visual comparison comes first.

plt.figure(figsize=(9,4))
plt.plot(test, label="Actual")
plt.plot(forecast, label="Forecast")
plt.legend()
plt.show()

What we see:

• Actual orders keep changing
• Forecast stays flat

This tells us the model is too simple.

Step 5: Residual Analysis

Residuals show what the model failed to capture.

residuals = test - forecast

plt.figure(figsize=(9,4))
plt.plot(residuals)
plt.title("Residuals")
plt.show()

Residuals are not random.

They still contain:

• Trend
• Seasonality

That means the model is underfitting.

Step 6: Improve the Pipeline

Now we know what to fix.

Next steps usually include:

• Removing trend (differencing)
• Modeling seasonality (SARIMA, features)
• Using ML or deep learning models

The pipeline repeats:

Build → Visualize → Evaluate → Improve

Why Pipelines Matter

Without a pipeline:

• Models are built randomly
• Evaluation is unreliable
• Production failures are common

With a pipeline:

• Decisions are systematic
• Errors are visible early
• Models improve logically

Practice Questions

Key Takeaways

• Forecasting is a process, not a single model
• Visualization guides every decision
• Evaluation tells what to improve next
• Strong pipelines lead to reliable forecasts

From the next lesson onward, we move into stronger classical forecasting models.