Hi @dennisbader — gentle nudge on this when you get a chance. CI hasn't triggered yet (looks like it needs a maintainer approve-and-run for first-time contributors). It extends the #3059 zero_division pattern to the percentage/range metrics (wmape, ope, arre, marre, coefficient_of_variation), with tests. Thanks!

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.48%. Comparing base (b0ccd3a) to head (d55581a).
⚠️ Report is 8 commits behind head on master.

@@            Coverage Diff             @@
##           master    #3122      +/-   ##
==========================================
- Coverage   96.55%   96.48%   -0.07%     
==========================================
  Files         160      160              
  Lines       17285    17293       +8     
==========================================
- Hits        16689    16686       -3     
- Misses        596      607      +11     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

Hey @mahimn01, thanks for this PR and contributing :). Will review it shortly as this is something we should definitely add, but it would be also great if you could create an issue for this first. Thanks!

Thanks @jakubchlapek, this makes sense — best-score for 0/0 is the right call, and it actually fixes a latent bug I'd missed: with NaN-everywhere, a genuinely perfect forecast on a degenerate series scores NaN (ope on a zero-sum series with an identical prediction; arre/marre on a constant series predicted perfectly).

I'll collapse _safe_pct_divide into _safe_scaled_divide — one helper with a configurable 0/0 fill (1.0 scaled, 0.0 percentage), x/0 → NaN, and "raise" preserved (keeping each caller's existing message). A nice side effect of the shared helper's element-wise fill: for arre/marre it fixes a case the current code gets wrong — a constant-actual column is currently all-NaN even at timesteps where the per-step error is 0; element-wise gives 0/0 → 0 and x/0 → NaN within the same column.

On scope (full breakdown in the issue): wmape/ope/arre/marre/coefficient_of_variation are already in this PR — I'll switch them to the unified helper. Good news on the rest of the family: sape/smape already encode 0/0 → 0.0 inline and have no x/0 case, so the only delta there is exposing the explicit zero_division knob for the "raise" option. ape/mape are the trickier ones — they still hard-raise on any y_t == 0, and because their denominator is the per-timestep actual (not a column scalar) they need element-wise handling; relaxing MAPE's default is also a larger behavioral change. Would you prefer ape/mape in this PR (default kept as raise, opt into fill via zero_division) or as a focused follow-up?

One small flag: coefficient_of_variation is RMSE/mean (not abs-wrapped, and the mean can be negative), so a zero mean is a legitimate non-degenerate case — I've proposed 0/0 → 0.0 / x/0 → NaN for it but happy to adjust.

Inline replies:

• (arre, "why removed?") that was the old if not (y_max > y_min).all(): raise(...) guard — the zero-range check moved into the helper, gated behind zero_division (default fills; "raise" re-raises, preserving the legacy behavior). The old message also said "MARRE" inside arre(), so that copy-paste slip drops out.
• (test) agreed — I'll fold the negative-sum check into test_ope() with a small negative-series helper. I'll also add a 0/0 best-score case to the zero-division test (the current fixtures are all x/0, so they stay NaN and need no change).

I've opened #3132 to track the design. Thanks for the thoughtful review!

Hey @jakubchlapek — no rush, just a bump on this one whenever you get a chance.

I think we're aligned on the core: fold the percentage path into the unified _safe_scaled_divide with best-score for 0/0 and x/0 → NaN. I opened #3132 to track the design.

Two calls are yours before I finish the rework:

1. ape/mape — happy to fold them into this PR or split them into a focused follow-up. They're a bit trickier than the rest since their denominator is per-timestep (so element-wise handling), and relaxing MAPE's default is a slightly bigger behavioural change — so I'd rather defer to you on scope.
2. coefficient_of_variation's 0/0 fill — I'm leaning 0.0 (RMSE = 0 is a genuinely perfect forecast), flagging it since its denominator can go negative.

Point me either way and I'll turn it around quickly, and rebase at the same time since master's moved.

Hey @mahimn01, thanks for the update.
Regarding the cv 0/0 fill it definitely makes sense for me to use 0.0 Maybe we could extend the _safe_scaled_divide to accept a best_score/zero_fill/fitting_name parameter. to pass for a metric what should be the output in case of 0/0.

Regarding the ape/mape I would definitely include them in this PR. This being said I am not quite sure what exactly the issue with them is for this approach. Similarly to other metrics (like coefficient_of_variation which you already implemented) ape exposes y_true and y_pred which we would just need to similarly wrap with the division helper. Similarly I don't get what you mean with the relaxing MAPE's default. Could you double check this and clarify? Thanks :)

Hey @jakubchlapek, agreed on both.

On ape/mape, you're right. Once _safe_pct_divide is folded into _safe_scaled_divide (which fills element-wise) it falls out with no special handling. You pass errors = y_true - y_pred and scale = y_true, both (t, c), and per element 0/0 gives 0.0 for a perfect forecast and x/0 gives NaN, same as cv. I'll include them.

On MAPE's default, ape and mape currently raise on a zero actual (it's documented), so defaulting zero_division="warn" changes that to warn-and-fill. Same raise→warn shift the PR already makes for the other five, so it stays under the 🔴 note. I'll default them to "warn" and update the docstrings.

I'll unify _safe_pct_divide into _safe_scaled_divide with a best_score arg (1.0 scaled, 0.0 percentage), route wmape, ope, arre, marre, cv, ape and mape through it, and rebase since master moved.

Reworked and pushed, also rebased onto master. _safe_pct_divide is now folded into _safe_scaled_divide via the best_score arg (1.0 scaled, 0.0 percentage), and ape, mape, wmape, ope, arre, marre and cv all route through it. 0/0 gives the perfect-forecast score and x/0 gives NaN, filled element-wise so a constant actual_series no longer NaNs a whole component.

One heads-up, since mape is the default metric for backtest and gridsearch, the raise to warn change makes those surface NaN instead of raising on a zero actual, consistent with the #3059 scaled metrics.