Rows that aren't independent
Almost every supervised learning technique you've used elsewhere on this site leans on one quiet assumption: the rows in your dataset are independent and identically distributed. Independent means knowing one row's label tells you nothing about another's. Identically distributed means every row was drawn from the same underlying process. Under that assumption, shuffling your data before splitting it into train and test sets is harmless, and in fact standard practice, because order carries no information.
Time series data violates both halves of that assumption. Today's value is correlated with yesterday's, which is correlated with the day before, often for reasons that persist for weeks or months (a heat wave, a marketing campaign, a supply shortage). And the process generating the data can itself change over time, so a model trained on last year's pattern may face a genuinely different distribution this year. Shuffle a time series before splitting it and you don't just lose information, you actively destroy the one thing that makes it a time series: the sequence itself.