Hydrological forecasting plays a critical role in water resources management and public safety. Although physically based models (PBM) have been widely used for hydrological forecasting, their applications in data-sparse high-latitude regions presents significant challenges. Deep learning methods such as long short-term memory (LSTM) integrating lagged observations have shown promises for streamflow prediction, but their effectiveness for forecasting remains uncertain. This study compared the performance of an LSTM model to that of a semi-distributed PBM, HYDROTEL, trained with the same historical data and forced with the same meteorological ensemble forecasts, to forecast short-range (0-14 days lead) streamflow and inflow at four sites in Yukon Territory, Canada. The LSTM model used Data Integration (DI) of recent observed flow and inflow, with or without DI of Snow Water Equivalent (SWE), while HYDROTEL used the same flow/inflow data as part of a Data Assimilation (DA)an Ensemble Kalman Filtering (EnKF) framework. Results demonstrated that for all lead times, the LSTM-DI model outperformed that of HYDROTEL-EnKF at three sites. When incorporating SWE data, the LSTM-DI model further surpassed HYDROTEL-EnKF at all sites. Compared to the version integrating only flow/inflow data, integrating SWE data offered meaningful advantages, particularly for predicting spring peak inflows at two sites, while performing similarly for the other sites. This study highlights the reliability of LSTMs for short-range hydrological forecasting, and the added value of including snow information on the forecasting skill in data-sparse high-latitude regions.