[BUG] fix transformer chaining in regression pipeline

[ashnaaseth2325-oss]

Summary

Fixes a mismatch between Pipeline._fit and _transform in:

skpro/regression/compose/_pipeline.py

During fit, transformers are chained correctly:

X = t.fit_transform(X=X, y=y)

But in _transform, X was never updated inside the loop. Each transformer received the original X, and only the last output was returned.

So a pipeline:

T1 → T2 → Regressor

was trained on:

T2(T1(X_train))

but predicted on:

T2(X_test)

---

Impact

• Multi-transformer pipelines produced incorrect predictions.
• CV and backtesting results could be silently corrupted.
• No error was raised — shapes remained valid.
• Single-transformer pipelines were unaffected (hard to detect).

---

Fix

Add one line inside _transform:

X = Xt

This properly chains transformer outputs.

---

Result

• Fit and predict now use the same feature representation.
• Multi-step pipelines behave correctly.
• No API changes.

[ashnaaseth2325-oss]

Hello @fkiraly , @felipeangelimvieira
This PR fixes an inconsistency where _transform wasn’t chaining transformer outputs like _fit, causing multi-step pipelines to use incorrect features at predict time.
Added X = Xt inside the loop so prediction now mirrors training . Happy to adjust if needed.

[fkiraly]

Thanks - this is a quite serious bug, thank you for fixing.

From a readability perspective, I would prefer if, inside the loop, you have Xt = transformer.transform(Xt) instead of constantly renaming the variable.

[ashnaaseth2325-oss]

Hello @fkiraly,
Thanks for the suggestion.
I've updated the loop to use Xt = transformer.transform(Xt) for improved readability and consistency with fit.
All checks are now passing. Happy to adjust further if needed.

[fkiraly]

Looks good, thanks!

It would be great if you could add a test that ensures the bug is fixed: I would suggest, pick any deterministic estimator, and then chain two exponentiation transformers or similar. Then check that using two of the transformers is indeed the correct exponent of using one.

[ashnaaseth2325-oss]

Hello @fkiraly
Thanks for the suggestion! I’ve added a test chaining two FunctionTransformer(np.exp) steps with a deterministic estimator and checking that the result matches exp(exp(X)).
CI is green now. Let me know if anything else should be adjusted.

[fkiraly]

Thanks. I was thinking more along the lines of comparing two pipelines, but this also works.