docs(userguide): clarify mc_dropout + likelihood combination

[jbbqqf]

Fixes #2105.

Summary

The User Guide section Capturing model uncertainty using Monte Carlo Dropout says

Monte Carlo Dropout can be combined with other likelihood estimation in Darts, which can be interpreted as a way to capture both epistemic and aleatoric uncertainty.

…but it never says how. As @SaltedfishLZX reported in #2105, this leaves users trying mc_dropout=True together with predict_likelihood_parameters=True and getting a single deterministic output, with no warning that the two flags shouldn't be combined.

This PR adds three things to that section:

1. The exact knob settings to obtain both epistemic and aleatoric uncertainty: train with likelihood=..., predict with mc_dropout=True, num_samples >> 1, predict_likelihood_parameters=False.
2. A one-line explanation of what each Monte Carlo pass actually does in that mode (one set of distribution parameters per pass, then one sample drawn per pass — so the resulting num_samples predictions reflect both sources of uncertainty).
3. A > Note flagging that mc_dropout=True + predict_likelihood_parameters=True should not be combined, with a link back to mc_dropout with predict_likelihood_parameters #2105 for context.

No runtime code is touched — only docs/userguide/forecasting_overview.md and a CHANGELOG entry.

Other Information

• Maintainer @dennisbader's response in mc_dropout with predict_likelihood_parameters #2105 already contained the substance of this clarification; this PR just lifts that explanation into the user guide so it surfaces before users hit the issue.
• The function-level docstring of predict() in darts/models/forecasting/torch_forecasting_model.py already states num_samples = 1 is required when predict_likelihood_parameters=True, so users do get a runtime error for the unsupported combination — no behaviour change is needed.
• CHANGELOG entry under Unreleased > Improved.

Reproduce BEFORE/AFTER yourself (copy-paste)

# --- one-time setup ---
git clone https://github.com/unit8co/darts.git /tmp/darts-2105 && cd /tmp/darts-2105

# --- BEFORE (origin/master): one-line vague claim ---
git checkout origin/master
sed -n '280,290p' docs/userguide/forecasting_overview.md
# Expected: a single sentence saying MC dropout "can be combined with
#           other likelihood estimation … both epistemic and aleatoric
#           uncertainty", with no concrete recipe.

# --- AFTER (this PR): concrete recipe + caveat ---
git fetch https://github.com/jbbqqf/darts.git docs/2105-mc-dropout-clarify
git checkout FETCH_HEAD
sed -n '280,300p' docs/userguide/forecasting_overview.md
# Expected: same first sentence, plus a bullet list of the exact knob
#           settings (mc_dropout=True, num_samples>>1,
#           predict_likelihood_parameters=False) and a note telling
#           readers not to combine mc_dropout=True with
#           predict_likelihood_parameters=True.

The sed line is identical between BEFORE and AFTER; only the checked-out ref differs.

What I ran locally

• Manually re-read the modified section against the issue thread (mc_dropout with predict_likelihood_parameters #2105) to confirm the recipe matches @dennisbader's explanation.
• Cross-checked darts/models/forecasting/torch_forecasting_model.py:1734-1736 to confirm the docstring already enforces num_samples = 1 for predict_likelihood_parameters=True, so the new note in the guide is consistent with current behavior.
• No code changes → no test changes; nothing in darts/tests/ exercises markdown.

Edge cases tested

#	Reader scenario	Before this PR	After this PR
1	Wants both epistemic + aleatoric uncertainty	Knows it's possible but not how	Has the exact knob settings inline
2	Tries mc_dropout=True + predict_likelihood_parameters=True	Surprised by deterministic output	Warned in the same section that this combination isn't supported
3	Already using mc_dropout=True with deterministic loss	Behaviour unchanged	Behaviour unchanged

Risk / blast radius

Documentation-only. No code, no APIs, no tests touched. Worst-case impact is a Sphinx build that picks up the new bullet list / blockquote — both standard Markdown.

---

PR drafted with assistance from Claude Code (Anthropic). The clarification was reviewed manually against the discussion in #2105 and the docstring of TorchForecastingModel.predict(). The reviewer can paste the BEFORE/AFTER block to verify.

[jakubchlapek]

Hey @jbbqqf, thanks for this PR!
Looks good for me, small clarification changes :)

[jakubchlapek]

feel free to take credit here (by ...) and refer to the PR number, not the issue

Suggested change

	- Clarified the `Capturing model uncertainty using Monte Carlo Dropout` section of the user guide to explain how `mc_dropout=True` interacts with the `likelihood`/`predict_likelihood_parameters` knobs, so readers know which combination produces both epistemic and aleatoric uncertainty. Closes [#2105](https://github.com/unit8co/darts/issues/2105).
	- Clarified the `Capturing model uncertainty using Monte Carlo Dropout` section of the `forecasting_overview.md` user guide to explain how `mc_dropout=True` interacts with the `likelihood`/`predict_likelihood_parameters` knobs. Closes [#3117](https://github.com/unit8co/darts/pull/3117).

[jakubchlapek]

small clarification

Suggested change

	> Note: `mc_dropout=True` and `predict_likelihood_parameters=True` should not be combined. `predict_likelihood_parameters=True` returns the *deterministic* parameters of the predicted distribution and bypasses the Monte Carlo sampling loop, so dropout has no observable effect on the output. See [issue #2105](https://github.com/unit8co/darts/issues/2105) for context.
	> Note: `mc_dropout=True` and `predict_likelihood_parameters=True` should not be combined. `predict_likelihood_parameters=True` is designed for use with `num_samples=1` to get the model's best estimate of the distribution parameters.