Break them down into multiple smaller formulas, with a column above explaining what they each do. Then consume the previous result in the next formula. This doesn't even need to be on the same sheet as the actual primary consumer sheet.

Break them into components and calculate the result iteratively if you can; I prioritize clarity and provability over raw performance. If it's a set of standard formula specific to an industry, consider lambda functions.

I use named ranges to semantically label relevant cells or ranges, and then use those labels in the formula. It's much easier to read and construct formulas when you don't have to think about where a value exists in the worksheet.

Named ranges! Also newer Excel versions provide you the LET function, which is neat for documentation (naming) of intermediate steps:

  =LET(data; F4:F34;
       dummy1; "This is a made up average function; this string is ignored";
       dummy2; "one limitation is that you cant really overwrite variable names";
       tally; COUNTA(data);
       total; SUM(data);
       total/tally)

It also avoids recomputing intermediate stuff. You can also encapsulate things in a LAMBDA anonymous function:

  =LET(data; F4:F34;
       mymeanfunction; LAMBDA(data;
                              LET(tally; COUNTA(data);
                                  total; SUM(data);
                                  total/tally));
       mean1; mymeanfunction(data);
       mean2; mymeanfunction(OFFSET(data;0;1));doc;"FIXME, this should have its own variable name to refer explicitly to G4:G34";
       AVERAGE(mean1;mean2))

You could also move the LAMBDA to the name registry and use that function similarly to how you would use a named range or a built-in formula!

Coming from a lisp background, I was ecstatic to see this, but they have heavy technical limitations. I did play a little bit though with these concepts and the dynamic array functions. Fun functions to explore:

  - DROP, TAKE, CHOOSECOLS, CHOOSEROWS to cut arrays
  - INDEX, OFFSET, COLUMN, ROW to navigate
  - WRAPCOLS, WRAPROWS, TOROW, TOCOL, VSTACK, HSTACK, TRANSPOSE to shape arrays
  - MAP, SCAN, REDUCE to compute on top of arrays

There's more!

  - Recursion is severely limited in lambda.
  - There are extra pains when putting formulas on the name manager or on conditional formatting (although more focused on references than on lambdas in particular, usage of lambda would be more powerful if they didn't exist)
  - You can't do arrays of arrays, so your data structures need to be flat

Depending on how many levels of indirection there are (i.e. how many formulas reference other cells/the results of other formulas in other cells) I've found that using "trace precedents" and "trace dependents" in the formulas tabe to be helpful.