Previous studies have shown the value of short-term (<6 hr) weather forecasts made by statistical equations derived solely from weather observations (an “obs-based” system). As technology and computing speed have advanced, the frequency of observations has become finer, both in time and space. In fact, the most recent obs-based systems are able to make skillful predictions, when compared to simple persistence, for lead times <60 min. Yet, it is hypothesized that the utility of an obs-based system begins to decrease for forecasts only minutes in advance. The system's inherent limitation stems from recognizing that highly accurate ultra-short-term forecasts can simply be made by repeating the current observation. Since persistence is already a powerful predictor, this reduces the potential benefit of supplemental observations as predictors and therefore the strength of an obs-based system. Pilot experiments testing this hypothesis have been conducted using an archive of AWS/WeatherBug observational surface data of one-minute temporal frequency. Preliminary results on an independent sample, comprising 70 cases, indicate that obs-based temperature forecasts for Ardmore, Pennsylvania, no longer are statistically superior (using α = 99%) to corresponding persistence forecasts for lead times shorter than ~20 min. Additional results will be presented for other weather parameters (e.g., dew point, wind speed) and other AWS sites. Although these results have theoretical merit in quantitatively assessing atmospheric predictability for the ultra-short-term, this study also has practical application, namely serving to optimize future statistical forecasting products.