In this case, a cesium-beam frequency standard (aka atomic clock.) The specific data source isn't as important as the noise processes that it exhibits -- in this case a combination of white and random-walk frequency noise. The latter noise type alters the phase slope over multiple timeframes at once, even though the slope is accurately known over the long term.
The usual metaphor is a drunk person looking for his lost car keys. He meanders around under the streetlight because that's the only place where he can see where he's going. He won't stray very far from the lamp post, but his direction at any given time has little or no correlation to either his past or future behavior.
It's easy to fool yourself into thinking you understand what's going on based on recent historical behavior, but in reality, the presence of random-walk noise means that it's impossible to infer anything about long-term trends or short-term biases by looking at short-term trends. In climate science, even a hundred thousand years' worth of data is still a "short term" record. We need better data, we need better models, and most important, we need to give ourselves time to evaluate them on the basis of their predictive power.
Based on my own experience watching random-walk processes in real time, someone who expects me to take action based on the last 100 years of data from a multi-billion year timeframe is just going to get laughed at. I've spent so much time fooling myself (that's my software, and my cesium standard) that I probably am erring on the side of too much skepticism.