ARIMA Forecasting

Automatic ARIMA/SARIMA — classic statistical univariate time series forecasting

---

Overview

• Model ID: arima-forecasting
• Framework: pmdarima (auto_arima)
• Best for: Univariate time series, interpretable statistical models
• GPU required: No (CPU-only)
• Training time: 1-10 minutes

ARIMA (AutoRegressive Integrated Moving Average) is a classic statistical model for time series forecasting. ColabHive uses pmdarima with automatic order selection (p, d, q) and optional seasonal components (P, D, Q, m).

---

When to Use

✅ Perfect for:

• Univariate time series (single series)
• Datasets with clear seasonal patterns (daily, weekly, monthly)
• Use cases requiring interpretable models
• Small datasets (ARIMA works well with few data points)
• Production systems requiring fast, lightweight inference

❌ Not ideal for:

• Multi-column forecasting (use classical-forecasting-auto)
• Very long-horizon forecasts (>1 year) with complex patterns
• Non-stationary series without differencing

---

Quick Start

from colabhive import ColabHive

client = ColabHive(api_key="...", account_id="...")

# Upload time series dataset
dataset = client.datasets.upload(
    name="sales-series",
    file="./sales.csv",
)

# Train ARIMA (auto order selection)
job = client.training.create(
    model="arima-forecasting",
    dataset_id=dataset.id,
    hyperparameters={
        "date_column": "date",
        "target_column": "sales",
        "seasonal": True,
        "m": 12,  # Monthly seasonality
    }
)

job.wait()
print(job.get_metrics())

---

Dataset Format

Your CSV should have at minimum a date column and a target column:

date,sales
2023-01-01,1200
2023-02-01,1350
2023-03-01,1100
...

Requirements:

• Date column must be parseable by pandas (YYYY-MM-DD recommended)
• Minimum 24 data points for reliable order selection
• Regularly spaced time intervals (daily, monthly, etc.)

---

Hyperparameters

Parameter	Default	Description
date_column	"date"	Name of the date/timestamp column
target_column	"target"	Name of the column to forecast
seasonal	true	Enable seasonal ARIMA (SARIMA)
m	1	Seasonal period (12=monthly, 7=daily-weekly, 4=quarterly)
max_p	5	Max AR order to search
max_q	5	Max MA order to search
max_d	2	Max differencing order
information_criterion	"aic"	Model selection criterion (aic, bic, hqic)
forecast_horizon	12	Number of periods to forecast

Common m values

Frequency	m value
Daily (weekly seasonality)	7
Monthly (yearly seasonality)	12
Quarterly	4
Weekly (yearly seasonality)	52
No seasonality	1

---

Output Metrics

After training, job.get_metrics() returns:

{
  "aic": 245.3,
  "bic": 260.1,
  "mae": 42.5,
  "rmse": 58.3,
  "mape": 3.8,
  "order": [2, 1, 1],
  "seasonal_order": [1, 1, 0, 12]
}

• AIC/BIC: Lower is better (model complexity penalty)
• MAE: Mean Absolute Error
• RMSE: Root Mean Squared Error
• MAPE: Mean Absolute Percentage Error (%)
• order: Final (p, d, q) selected
• seasonal_order: Final (P, D, Q, m) selected

---

Register and Use for Inference

# Register as inference endpoint
endpoint = client.training.register_for_inference(
    run_id=job.id,
    name="arima-sales-forecast",
    description="Monthly sales ARIMA model",
    visibility="account",
)

# Predict next 6 months
result = client.endpoints.infer(
    endpoint_id=endpoint.endpoint_id,
    input_data={"forecast_horizon": 6},
)
print(result["result"])

---

Comparison with Other Time Series Models

Model	Univariate	Multi-series	GPU	Auto-select	Best for
arima-forecasting	✅	❌	❌	✅	Classic statistical, interpretable
prophet-forecasting	✅	❌	❌	—	Seasonal patterns, holidays, credible intervals
classical-forecasting-auto	✅	✅	❌	✅	Multi-column, auto model selection per series
timesfm-2.5-finetune-gpu	✅	✅	✅	—	Foundation model, quantile predictions, complex patterns

---

Learn More

• Classical Forecasting (Auto) — multi-series auto-selection
• Prophet Forecasting — holiday/seasonality decomposition
• TimesFM 2.5 — GPU foundation model
• Preparing Datasets
• Hyperparameter Tuning