The accuracy of the predicted Haines Index using the MM5 model is evaluated using rawinsonde observations. The evaluation compared predicted and observed Haines Indices during a 5-month period from 1 June to 31 October 2003 at 29 upper-air sounding sites within the modeling domain over much of the eastern half of the USA. Despite a consistent cold bias of 1–3°C in the lower to mid troposphere and a dry bias at the 850-mb and 700-mb levels in the MM5 results, little bias is found in the predicted Haines Index. This is because the temperature bias is largely cancelled in the calculation of the vertical temperature gradients used to derive the Haines Index and the dry bias is compensated by the slightly weaker instability in the model predictions. The predicted Haines Index captured the observed spatial and temporal variations of the observed index values resulting from changes in synoptic conditions. Of the 4184 total available soundings for the comparison, 45% had an exact match between the predicted and observed Haines Index categories and another 43% had a good match with the predicted and observed index differing by only one category. The rest of the 12% were cases where the predicted index fell under a category that was substantially different from the observed category, significantly raising or lowering the risk level for potentially dangerous fire behavior. Despite the general success, serious limitations exist when using the Haines Index derived from MM5 results for predicting high-risk conditions. This is because the prediction failed more than 50% of the time to capture the observed extreme cases when the observed index reached its highest possible value of 6. Although such extreme cases are rare, they represent conditions that are most conducive to dangerous and erratic fire behaviors and failure to predict these conditions may lead to serious consequences.