We investigate the role of pre-eruptive magma conditions in controlling eruption styles at basaltic andesite volcanoes, and their relationship to monitoring data. We focus on a cycle of explosive (1990, sub-Plinian) => effusive (2007 dome) => explosive (2014, sub-Plinian) eruptions at Kelud volcano, Indonesia. We conducted petrological analyses of eruption products and phase equilibria experiments using a pumice from the 1990 eruption across a wide range of conditions (temperature, pressure, oxygen fugacity, and volatile contents). We are able to reproduce the main mineral assemblage of all three magmas (plagioclase + pyroxenes + amphibole + olivine + magnetite) at T = 975 ± 39°C, P = 175 ± 25 MPa, NNO, and 4-6 wt. % water in the melt, which is also consistent with the phenocryst contents for explosive eruptions (30-50 wt. %). The high phenocryst content of the 2007 dome (~ 70 wt. %) implies a lower volatile content (~1 wt. % water in the melt), which likely resulted from slow magma ascent towards the surface alongside progressive equilibration. Mass balance calculations on the sulfur budget of the 1990 and 2014 magmas, show that they contained an excess fluid phase that was stored with the magma, and we propose that this led to their more explosive eruption style compared to the 2007 dome. Our hypothesis of accumulation of excess fluids could perhaps be recorded as changes in gravity data over time, and may open doors to better anticipate eruption styles.