High frequency sensor data in rivers can greatly increase understanding of ecological processes, thus supporting management decisions. Such data allow testing and comparing multiple models of varying complexity. Here we evaluate this tradeoff between model simplicity and complexity using time series models of daily estimates of riverine photosynthesis and respiration. Simple models can adequately fit the data, but may not reflect the underlying processes, e.g., hydrologic scouring of biomass during floods. More complicated models allow partitioning error into observation and process error and modeling biomass as latent variables, but using simulated data we show how parameters in such models may not always be easily identifiable. An alternative approach is to build models where data define the model. We demonstrate a simple method for selecting the relevant covariates and time lags from a large number of possible predictors of a photosynthesis time series using regularization to account for sparsity. Such models allow exploring time scales of memory in rates of photosynthesis and may predict future states well. Combining these various time series approaches will advance or understanding of controls on metabolism and provide a basis for using metabolism as an indicator in ecosystem management.