it can not be due to variance??? Who's turning on new miners? New Chinese mining farm(s)? Don't they know btc price is in the dumps???
Don't underestimate the size of the daily variation in hash rate due to statistical noise. The hash rate regularly varies by 10% to 20% from day to day solely due to statistical variance. It is only over a week or longer that actual changes in the hash rate become clear.
You should assume a 10 to 20% error in the hash rate if data is taken over just one day.
If data is taken over a longer period the error due to statistical variance declines but it is still there.
In particular you can still expect the hash rate difficulty changes every two weeks to fluctuate within a range of about plus and minus 2-4% solely due to statistical noise even if there is no change to the total actual network hash rate.
The mathematical details.
There is a 10 minute block average solve time with a 10 minute variance (poisson distribution). So the standard deviation percentage error in solve time is 100 / sqrt(number blocks). Most of the time the difficulty change will be within 2 standard deviations.
There are 144 blocks found on average each day so for one day the standard deviation error is 100 / sqrt(144) or about 8%. Two deviations are 16%
There are 2016 blocks found on average between each difficulty change so for each change the standard deviation is 100 / sqrt(2016) or about 2%. Two deviations are 4%. So hash difficulty changes may be frequently in error by 2% and possibly as high as 4%.