The cortical learning algorithm (CLA) is a time-series forecast method designed based on the human neocortex. CLA has memory components named columns. A column synapse links a column to an input data bit. CLA encodes real value input data in each time step into a discrete representation based on column synapse relationships. CLA also represents the data incoming next time step as a discrete representation. A column-based decoder decodes the discrete representation into a real value based on column synapse relationships. Therefore, column synapse relationships have a large impact on forecast accuracy. In this work, we propose a CLA introducing an adaptive column synapse arrangement that improves the column synapse relationships for the column-based decoder. We showed that the adaptive column synapse arrangement boosts the forecast accuracy improvement of the column-based decoder on artificial time-series data and real-world electricity load forecast.