A code-modulated Visual Evoked Potential Brain Computer Interface (c-VEP BCI) allows for spelling from a virtual keyboard of flashing characters. All characters flash simultaneously, and each character flashes according to a predefined pseudo-random binary sequence, circular-shifted by a different time lag. For a given character, the pseudo-random stimulus sequence evokes a VEP in the electroencephalogram (EEG) of the subject, which can be used as a template. This template is usually obtained during a calibration phase and it is applied for the target identification during the spelling phase. A downside of a c-VEP BCI system is that it needs a long calibration phase to reach good performance. This paper proposes an unsupervised method that avoids the calibration phase in a c-VEP BCI, by extracting relative lags from the VEP responses, between successive characters, and predicting the full word using a dictionary. We tested it in offline experiments on a public dataset. We simulated the spelling of four groups of words with a different total number of characters selected from an English dictionary. Each experiment is parameterized by the number of stimulus cycles. The obtained results show that a word-prediction-based auto-calibration method in c-VEP BCIs can be efficient and effective.