Forecast

📈 What is Amazon Forecast?

Amazon Forecast is a fully managed service that uses machine learning to deliver highly accurate time series forecasts. It is based on the same technology Amazon.com uses for predicting demand, sales, and inventory.

✅ No ML experience required. Just bring historical time-series data and related metadata (like item ID, category, location).

---

📦 Use Cases

Domain	Example Forecasting Use Case
Retail	Demand forecasting, product sales
Supply Chain	Inventory planning, stockouts
Energy	Power demand prediction
Finance	Revenue projection, cost estimation
Healthcare	Bed availability, patient intake
Workforce	Staff scheduling, call center volume

---

🚀 Forecast Capabilities

Feature	Description
AutoML	Selects the best algorithm automatically
Domain-specific models	Includes CNN-QR, DeepAR+, Prophet, ARIMA
Related time series	Incorporate external data like weather, price, holiday
Item metadata support	Use product category, brand, etc. to improve accuracy
Probabilistic forecasts	Predict confidence intervals (P10, P50, P90)
Retrainable & reusable models	Models can be retrained and reused
Export to S3 or inference	Run batch forecasts or export CSV to S3

---

📁 Required Input Datasets

Dataset Name	Required	Description
Target Time Series	✅ Yes	Historical demand per item per timestamp
Related Time Series	Optional	Weather, promotions, price, etc.
Item Metadata	Optional	Product category, brand, warehouse, etc.

✅ Must include:

• item_id

• timestamp

• target_value

---

🧠 Forecasting Workflow

1. Prepare Data → Upload CSVs to S3

2. Create Dataset Group

3. Create Predictor (train model)

4. Create Forecast

5. Query or Export Forecast Results

---

🧪 Sample Python (Boto3) Code

Step 1: Create Dataset Group

import boto3
forecast = boto3.client('forecast')

forecast.create_dataset_group(
    DatasetGroupName='sales-forecast-group',
    Domain='RETAIL',
    DatasetArns=[
        'arn:aws:forecast:...:dataset/target_time_series'
    ]
)

---

Step 2: Train Predictor

forecast.create_predictor(
    PredictorName='sales-predictor',
    ForecastHorizon=30,
    PerformAutoML=True,
    InputDataConfig={
        'DatasetGroupArn': 'arn:aws:forecast:...:dataset-group/sales-forecast-group'
    },
    FeaturizationConfig={
        'ForecastFrequency': 'D',
        'Featurizations': [{'AttributeName': 'target_value', 'FeaturizationPipeline': []}]
    }
)

---

Step 3: Generate Forecast

forecast.create_forecast(
    ForecastName='sales-forecast',
    PredictorArn='arn:aws:forecast:...:predictor/sales-predictor'
)

---

Step 4: Query Forecast Results

query = boto3.client('forecastquery')

response = query.query_forecast(
    ForecastArn='arn:aws:forecast:...:forecast/sales-forecast',
    StartDate='2025-07-01T00:00:00',
    EndDate='2025-07-31T00:00:00',
    Filters={"item_id": "item-001"}
)

print(response['Forecast']['Predictions'])

---

🧾 Output Example

{
  "Predictions": {
    "p10": [{"Timestamp": "2025-07-01", "Value": "125.5"}, ...],
    "p50": [{"Timestamp": "2025-07-01", "Value": "140.7"}, ...],
    "p90": [{"Timestamp": "2025-07-01", "Value": "160.2"}, ...]
  }
}

---

💰 Pricing (2024)

Component	Price
Training (AutoML)	$0.60 per training hour
Inference	$0.00075 per forecasted data point (per item, per timestamp)
S3 Storage	Standard S3 pricing
Free Tier	10,000 forecast units/month for first 2 months

🧠 Forecast Unit = 1 item × 1 timestamp.

---

🛡️ Security

Feature	Support
IAM role for access	✅ Required for S3 + Forecast
KMS Encryption	✅ Optional for input/output
VPC Support	❌ Not supported
CloudTrail	✅ Yes for API logging

---

🧱 Terraform Example

Amazon Forecast has limited Terraform support. Here's how to create a dataset:

resource "aws_forecast_dataset" "target_dataset" {
  dataset_name = "my-forecast-dataset"
  domain       = "RETAIL"
  dataset_type = "TARGET_TIME_SERIES"
  data_frequency = "D"
  schema = jsonencode({
    Attributes = [
      { AttributeName = "item_id", AttributeType = "string" },
      { AttributeName = "timestamp", AttributeType = "timestamp" },
      { AttributeName = "target_value", AttributeType = "float" },
    ]
  })
}

---

🧠 Model Types Used by Forecast

• DeepAR+ – probabilistic RNN (great for sparse and intermittent demand)

• CNN-QR – convolutional quantile regression

• ARIMA – traditional time series

• Prophet – from Facebook, seasonal forecasting

• ETS – exponential smoothing

Forecast automatically selects the best model using AutoML (unless overridden).

---

✅ TL;DR Summary

Feature	Amazon Forecast
Forecast type	Time series (quantitative)
Manual or AutoML	✅ Both supported
Forecast horizon	Up to 500 days
Confidence intervals	✅ P10, P50, P90
Input format	CSV (uploaded to S3)
Inference options	Batch queries via API or S3 export
Free Tier	✅ 10K forecasts/month (2 months)
Terraform support	✅ Partial (dataset/dataset group only)
Common integrations	S3, Lambda, Athena, Step Functions

---