We study the consistency of long-term trends in the surface water balance of 63 sub-catchments of the Amazon River basin. Monthly time series of precipitation, evaporation, runoff and soil water storage are obtained from remote sensors and flow stations (CHIRPS, ETR-Amazon, ANA-Brazil and JPL-GLDAS). Missing data during the period 1995-2015 are reconstructed applying an adaptation of the methodology proposed by Kondrashov & Ghil (2006). Empirical mode decomposition (Huang et al., 1998) is applied to filter out different modes of natural variability, with the aim to isolate the long-term trend of time series. The Mann-Kendall and Sen tests are applied to the residue and the sign and magnitude of the trends was obtained. No generalized unidirectional trends were found for the Amazon basin for any of the variables studied (Table 1), although some unidirectional trends were found in groups of sub-catchments that belong to the same stream. The consistency of the general water balance equation [dS/dt=P(t)-E(t)-R(t)], and its long-term approximation [\overline{R} = \overline{P} - \overline{E}] were evaluated. The general water balance does not close for 37 sub-catchments, while in the long-term the error in the balance tends asymptotically to a constant value, different from zero, which indicates that in the period of 20 years studied the long-term condition is fulfilled, but there is no closure for the long-term water balance either. The consistency of the surface water balance equation was also studied regarding the signs of the trends, finding that in 32 (51%) of sub-catchments studied the trend signs are not consistent with the water balance equation. Finally, for the remaining 31 sub-catchments (with trends consistent in signs), the consistency of the water balance equation was evaluated regarding the magnitude of trends, finding closure errors on water balance of trends up to 281% of the average of the magnitudes. We discuss several reasons for the lack of closure of the water balance, including the very existence of trends that violate the stationary hypothesis underlying the long-term approximation of the surface water balance. The used data and some of the results are available on the supplemental files. The Raw Data CSV file contains the spatial average timeseries at the basin that drain to the flow stations for the water balance equation variables. The Filled Data CSV file contains those timeseries after appliying the reconstruction. MKtrends and SenTrendMagnitude CSV files contains the results of the appliying the Mann-Kendall and Sen trend tests to the filled data.