1. Definition

A naïve forecast is a baseline model where the forecast for the next time step is simply the last observed value.

$\hat{y}_{t+1} = y_t$

  • “Tomorrow = Today” rule.
  • It assumes the most recent observation is the best predictor of the future.

2. Variants

  • Naïve 1 (Random Walk forecast):
    • $\hat{y}_{t+1} = y_t$
    • Commonly used in financial time series (stock prices).
  • Seasonal Naïve forecast:
    • Forecast = value from the same season in the previous cycle.
    • $\hat{y}_{t+h} = y_{t+h-m}$​ where $m$ = season length (e.g., 12 for monthly data).
    • Example: Forecast July 2025 sales using July 2024 sales.

3. Why It’s Important

  • Baseline benchmark: Any sophisticated forecasting model (ARIMA, LSTM, Prophet, etc.) should beat the naïve forecast.
  • M-Competitions (Makridakis competitions): showed naïve and exponential smoothing methods are often surprisingly competitive.
  • Provides a simple sanity check — if your advanced model can’t outperform the naïve baseline, it may be overcomplicated or mis-specified.

4. Example

Suppose monthly sales (units):

  • Jan: 100
  • Feb: 120
  • Mar: 130

Naïve forecast for April:

$\hat{y}_{April} = y_{March} = 130$

Seasonal naïve (if yearly seasonality, $m=12$):

$\hat{y}_{April 2025} = y_{April 2024}$

5. Evaluation

When evaluating forecasting models, the naïve forecast is often used as a baseline:

  • Metrics like MASE (Mean Absolute Scaled Error) explicitly scale forecast error relative to the naïve baseline.
  • If MASE < 1 → model beats the naïve forecast.

6. Applications

  • Finance: Random walk models of stock prices often perform as well as complex models.
  • Retail/Energy: Seasonal naïve forecasts are simple but effective for demand/load with strong seasonality.
  • Competitions: Always included as a benchmark.

Summary:
A naïve baseline forecast predicts the future using the most recent observation (or same season’s past value in seasonal cases). It’s simple, transparent, and often competitive, making it an essential benchmark that every forecasting model should outperform.