Small forecasting fixes: forecast-only future regressors and dict-typed CLI options

[BelhsanHmida]

Description

This PR fixes two independent bugs found while testing the forecasting pipeline with ENTSO-E day-ahead fundamentals (residual load, load forecast, generation outages) as future regressors. Both are small and come with regression tests.

• Forecast-only future regressors reach the training window: sensors that only ever record forecasts (belief time never after the event start, e.g. day-ahead market fundamentals) were filtered out of the training window entirely, so the model trained on a near-constant, interpolation-filled column and silently ignored the regressor
• Realized values still win: where a realized belief exists for an event, it takes precedence over the forecast, per event and per regressor
• Strict semantics preserved: past regressors keep their realized-only contract, and the predict-window slice stays forecast-only, so no ex-post data leaks into a simulated backtest
• Dict-typed CLI options parse again: plain fields.Dict options such as --model-params (and --snap) were passed to Marshmallow as raw strings and rejected with "Not a valid mapping type"; they now reuse NestedDictParamType, accepting both JSON and Python-literal syntax
• Test coverage for the Click layer: the dict-option bug shipped unnoticed because existing tests construct the model directly and never exercise the CLI, so the new test invokes the option through CliRunner

Why the regressors went missing

split_data_all_beliefs built the training portion of the future-regressor covariates by keeping only realized beliefs (belief_time > event_start). A day-ahead fundamental is always recorded ~21h before its event, so it has no such beliefs: every training row was dropped, and the missing-value filler papered over the gap by stretching a single value across it.

The effect was total. In my testing, 91% of the covariate series that reached the model was one repeated value, injecting an absurd synthetic signal changed none of the 672 predictions, and forecasting with --future-regressors was bit-identical to a run with no regressors at all.

_latest_known_per_regressor now takes a fall_back_to_forecast flag. When set, it keeps both ex-post and ex-ante beliefs and lets the existing "latest belief per event" selection decide: for any given event a realized belief time necessarily exceeds every forecast belief time, so realized values win wherever they exist and other events fall back to their latest forecast. Only the training-window call site passes the flag.

Example

Both of these now work as documented:

# A future regressor whose sensor only ever records day-ahead forecasts
# now populates the training window instead of collapsing to a constant.
flexmeasures add forecasts --sensor 14 --future-regressors 796 \
    --train-start 2026-03-01T00:00:00+01:00 \
    --start 2026-05-11T00:00:00+02:00 --end 2026-05-18T00:00:00+02:00

# Dict-typed options are parsed at the Click level (JSON or Python-literal).
flexmeasures add forecasts --sensor 14 --model-params '{"max_depth": 6}'

Before this PR the second command failed with:

Error: Invalid forecasting configuration: {'model-params': ['Not a valid mapping type.']}

Testing Instructions

Run the two affected test modules:

pytest flexmeasures/data/tests/test_forecasting_pipeline.py
pytest flexmeasures/cli/tests/test_utils.py

• test_forecast_only_future_regressor_populates_training_window covers a future regressor with only ex-ante beliefs, plus a second regressor with one realized revision to prove realized values still take precedence per event.
• test_add_cli_options_from_schema_parses_dict_fields invokes --model-params through Click in both JSON and Python-literal syntax and checks the parsed value passes schema validation.
• Both tests fail on the parent commit and pass with the fix. Full modules green: 50 passed and 11 passed / 1 xfailed respectively.

Verified against real data too (NL day-ahead prices with ENTSO-E residual load as a future regressor): before the fix the training covariate was 91% a single constant; after it, all 6,812 training values vary as they should.

Sign-off

• I agree to contribute to the project under Apache 2 License.
• To the best of my knowledge, the proposed patch is not based on code under GPL or other license that is incompatible with FlexMeasures

[read-the-docs-community]

Documentation build overview

📚 flexmeasures | 🛠️ Build #33844254 | 📁 Comparing 689b800 against latest (26d71eb)

  🔍 Preview build  

23 files changed · ± 20 modified · - 3 deleted

± Modified

• changelog.html
• genindex.html
• py-modindex.html
• _autodoc/GROUP.html
• _autodoc/INFLEXIBLE_DEVICE_SENSORS.html
• _autosummary/flexmeasures.api.v3_0.assets.html
• _autosummary/flexmeasures.api.v3_0.sensors.html
• _autosummary/flexmeasures.cli.utils.html
• _autosummary/flexmeasures.data.models.generic_assets.html
• _autosummary/flexmeasures.data.models.planning.devices.html
• and 10 more...

- Deleted

• _autodoc/INFLEXIBLE_CONSUMPTION.html
• _autodoc/INFLEXIBLE_PRODUCTION.html
• _autosummary/flexmeasures.data.schemas.scheduling.groups.html