RE: Which forecasting method works best with very little monthly demand data?

With only a small amount of monthly demand data (for example, 18–24 observations), simpler forecasting methods often outperform more complex machine learning models. The reason is that advanced models usually require much more historical data to learn meaningful patterns without overfitting.

Here are some approaches worth considering:

Simple Exponential Smoothing (SES): Best if your data has no clear trend or seasonality.
Holt’s Linear Trend Method: A good choice if demand shows a gradual upward or downward trend.
Holt-Winters Exponential Smoothing: Useful if you have enough data to capture seasonality (typically at least two full seasonal cycles).
ARIMA: Can perform well with small datasets, but requires careful parameter selection and validation.
Moving Average: Easy to implement, though it may lag behind when demand changes quickly.

I would generally avoid training machine learning models like Random Forests, XGBoost, or neural networks unless you have additional features (such as promotions, holidays, pricing, or weather) and a much larger dataset.

Model Validation

Since the dataset is small, avoid using a random train-test split. Instead, use time series cross-validation (also called rolling or walk-forward validation), where each forecast only uses past observations to predict future values. This provides a more realistic estimate of how the model will perform in practice.

Practical Recommendation

If I were starting with only 18–24 months of monthly demand data, I would:

Plot the data to identify trends or seasonality.
Build a simple baseline using a moving average or naïve forecast.
Compare it against Exponential Smoothing and ARIMA.
Evaluate the models using metrics such as MAE, RMSE, or MAPE.
Choose the simplest model that delivers consistent forecasting accuracy.

In many real-world demand forecasting problems with limited historical data, well-tuned statistical methods like Exponential Smoothing or ARIMA often provide more reliable forecasts than complex machine learning models. As more data becomes available, you can then consider more advanced approaches if needed.