Accepted answer
Four months is about 17 weeks, or 122 days — long enough that it is neither, and both answers are usually wrong in the same way. Take measurement error first, because it is cheap to exclude. Day-to-day scale variation of one to two kilograms is ordinary, so a stall is only visible in a rolling average; over 17 weekly averages a genuine flat line is unmistakable and a noisy plateau is not. If you have single weigh-ins rather than averages, you do not yet have a stall, you have a scatter plot. Then adaptive thermogenesis. It is real and it is smaller than the internet believes: metabolic rate falls with mass, and the residual adaptation beyond that is on the order of 50 to 100 kcal a day. Over 122 days that is 6,000 to 12,000 kcal, which is one to one and a half kilograms of tissue — enough to slow a curve, nowhere near enough to flatten it for four months. The arithmetic that closes the gap is intake drift. A 17-week plateau needs an energy balance of zero, and at this stage of loss that is a few hundred kilocalories a day of unrecorded intake — which is one of the best-replicated findings in the whole literature and the least popular. Weigh food for a fortnight rather than arguing about the mechanism. It settles the question in 14 days out of 122.
Answering this needs the actual intake data, because unmeasured intake drift is the largest single contributor and the hardest to see from inside.
Intake drift is real and invisible: portion sizes creep, untracked items accumulate, and self-reported intake is systematically underestimated even by people who are trying hard.
Water retention masks fat loss on a timescale of weeks. Glycogen repletion, sodium changes and menstrual cycle phase can each hide a kilogram or more of genuine loss.
Self-reported energy intake is underestimated by twenty per cent or more in doubly labelled water comparisons, consistently across populations.
Nothing here is medical or dietetic advice.
Your maintenance requirement fell because you did. Recalculate rather than escalate.
edited 6 Apr 2026 by m_haraldsen — added a caveat about sampling