If I have correctly understood your query; I think that you should check if the recorded temperature at the TGA plot is taken from the reference pan, rather than from the sample pan. If taken from the sample pan; the temperature can indeed retrocede, for a short time period, due to heat absorption by the sample during an endothermic process ― even if the temperature at the reference pan and at the furnace keeps increasing with elapsed time. Obviously, the temperature taken from the reference pan is not likely to retrocede while the furnace temperature increases; except for highly incorrect temperature control (PID) parameters, for some spurious power failure, or for some kind of instrument or software malfunction which would have to be carefully troubleshooted. Please check if the analysis heating schedule was correctly implemented, from well-below to well-above the investigated thermal process, to avoid possible instrumental artifacts, as they are more likely to occur by the extremes of the temperature ramp. It may help to check the TGA plot vs. elapsed time. Of course, the TGA instrument should have been adequately calibrated.

I can't think of any chemistry which might be causing problems, but have you run a blank? i.e. put in an empty crucible and just run the same program and see if you see the same "jump" back in temperature? It sounds more like a machine fault/machine sensitivity than chemistry. 

It's possible that at 550C, this is the point where there's a spike in current to the heater (the heater will be needing to "work" harder). If you're working with really small sample masses (and therefore have a noisy baseline) you might be detecting something like this... 

If I have correctly understood your query; I think that you should check if the recorded temperature at the TGA plot is taken from the reference pan, rather than from the sample pan. If taken from the sample pan; the temperature can indeed retrocede, for a short time period, due to heat absorption by the sample during an endothermic process ― even if the temperature at the reference pan and at the furnace keeps increasing with elapsed time. Obviously, the temperature taken from the reference pan is not likely to retrocede while the furnace temperature increases; except for highly incorrect temperature control (PID) parameters, for some spurious power failure, or for some kind of instrument or software malfunction which would have to be carefully troubleshooted. Please check if the analysis heating schedule was correctly implemented, from well-below to well-above the investigated thermal process, to avoid possible instrumental artifacts, as they are more likely to occur by the extremes of the temperature ramp. It may help to check the TGA plot vs. elapsed time. Of course, the TGA instrument should have been adequately calibrated.

i would also look for the temperature-time plot.

As you describe it there is a mass loss of 98%. Is the experiment run under nitrogen? or in air? in air the high exothermal heat of reaction can cause a fast increase in sample temperature (seen from temperature-time plot) which then drops when the reaction ceases.

Maybe you could share a plot here? It would be helpfull for interpretation.

Your sample undergoes a thermal runaway. You can avoid it reducing the sample mass. Above a critical mass heat released by the exothermic reaction exceeds heat dissipation. Then, the sample temperature locally rises and a combustion front propagates through the sample. Once the reaction is completed, the front extinguishes and the sample is again thermalized with the apparatus. I have a paper published that presents a criterion to avoid this overheating.

By the way, if the critical mass of your sample for thermal explosion to occur is too low you can try to mix your sample with an inert compound such as alumina powder. This may also work.