Subseasonal forecast of Arctic sea ice has received less attention than the seasonal counterpart, as prediction skill of dynamical models generally exhibits a significant drop in the extended range (>2weeks). The predictability of pan-Arctic sea ice concentration is evaluated by statistical models using weekly time series for the first time. Two statistical models, the vector auto-regressive model and the vector Markov model, are evaluated for predicting the 1979-2014 weekly Arctic sea ice concentration (SIC) anomalies at the subseasonal time scale, using combined information from the sea ice, atmosphere and ocean. The vector auto-regressive model is slightly inferior to the vector Markov model for the subseasonal forecast of Arctic SIC, as the latter captures more effectively the subseasonal transition of the underlying dynamics. The cross-validated forecast skill of the vector Markov model is found to be superior to both the anomaly persistence and damped anomaly persistence at lead times >3weeks. Surface air and ocean temperatures can be included to further improve the forecast skill for lead times >4weeks. The long-term trends in SIC due to global warming and its polar amplification contribute significantly to the subseasonal sea ice predictability in summer and fall. The vector Markov model shows much higher skill than the NCEP CFSv2 model for lead times of 3-6weeks, as evaluated for the period of 1999-2010.