A valid non-wear algorithm for activity monitors is crucial to avoid the misclassification of sedentary time as non-wear time, and vice versa. Characteristics of the algorithm, such as time windows, should be well defined and tested. Furthermore, using tri-axial data might influence the algorithm's performance. This study assessed the optimal time window length in a non-wear algorithm for overweight adults, applied to tri-axial data from sixteen participants. Ten time windows, from 10 up to 120 min, were tested with a diary as a criterion measure. We assessed the bias in non-wear time, sensitivity and specificity. The optimal time window length was based on ten participants; the validation of this time window was carried out with six other participants. The time window of 20 min showed the highest and 120 min showed the lowest mean amount of correctly classified non-wear time, at 94% and 70% respectively. Sensitivity and specificity were considered optimal in the 20 min time window. Validation of this time window demonstrated a sensitivity and specificity of 86% and 83% respectively. A 20 min time window showed the best non-wear estimations. The current study utilized tri-axial raw data and 1 s epoch data which might have facilitated the application of a short time window and thereby decreased the risk of misclassifying non-wear.