High crop yield variation between years, impacted for example by extreme weather shocks and by other shocks on the food production system, can have substantial effect on food production. This, in turn introduces vulnerabilities within global food system. To mitigate the effects of these shocks there is a clear need for understanding how different adaptive capacity measures link to the crop yield variability. While existing literature provides many local scale studies on this linkage, no comprehensive global assessment yet exists. We assessed reported crop yield variation for wheat, maize, soybean and rice for time period 1981-2009 by measuring both yield loss risk (variation in negative yield anomalies considering all years) and changes in yields during only dry shock and hot shock years. We used machine learning algorithm XGBoost to assess globally the explanatory power of selected gridded anthropogenic indicators (i.e., adaptive capacity measures; such as Human Development Index, irrigation infrastructure, fertilizer use) on yield variation on 0.5 degree resolution, within climatically similar regions to rule out the role of average climate conditions. We found that the anthropogenic indicators explained 40-60% of yield loss risk variation whereas the indicators provided noticeably lower (5-20%) explanatory power during shock years. On continental scale, especially in Europe and Africa the indicators explained high proportion of the yield loss risk variation (up to around 80%). Assessing crop production vulnerabilities on global scale provides supporting knowledge to target specific adaptation measures, thus contributing to global food security.