Abstract: Because fitted effect of threshold auto - regressive(TAR) model is sometimes better than its predicted effect, or predicted effect is bad, TAR is improved, i.e., when time series x(i) are f itted and predicted, ARs observed values of half period besides iare replaced by calculated values of fit or forecasting of TAR model. By depicting the f igure of auto correlation coefficients, it may ascertain clearly delayed paces a nd ranks of the auto - regressive model of threshold sections. The result of an example shows its improvement can enhance stability and practicability of the model and application of TAR in security supervision forecast is effective and successful. 