Monthly mean sink/source strength pattern of a degraded temperate peatland in northwestern Turkey was quantified using the eddy covariance technique between July 2010 and February 2014. Atmospheric and hydrological drivers and responses to extreme events of sink/source strength were analysed using multiple linear regression models, support vector machine, multilayer perceptron with principal component analysis and multiple comparisons following general linear models. On the basis of the three full-year measurements, mean annual, growing season and maximum estimates of net ecosystem CO2 exchange (NEE) were −0.22 mg CO2 m−2 s−1, −0.28 mg CO2 m−2 s−1 and −0.11 mg CO2 m−2 s−1 in July, respectively. The main driver of changes in monthly mean NEE pattern was found to be latent heat flux. Independent validation results showed a similar predictive performance (r2 = 74%) of multiple linear regression and support vector machine models as a function of month, total rainfall, water table depth, relative humidity, latent heat, atmospheric water vapour, wind direction and downwelling shortwave radiation. A shift in NEE from a net sink to a net source was observed in response to the deepest and shallowest water table depths due to drought and flooding, respectively. Copyright © 2014 John Wiley & Sons, Ltd.