[Anish Mehta (anishmehta24)]

Why are these changes needed?

add_time_idx_col (used for ts_forecast_panel when the data has no time_idx column, and in the TFT predict path) is broken for any frequency other than monthly/yearly:

• timestamps - timestamps.min() / freq only divides the minimum, so time_idx ends up as the raw nanosecond timestamp. For daily data starting 2024-01-01 the first value is 1704067199999980288 instead of 0.
• The values are computed from the unique dates and assigned back to X, so every row whose timestamp repeats (any panel with more than one series) gets NaN, and .astype("int") raises IntCastingNaNError.
• Series.view was removed in pandas 3, so there it fails with AttributeError before any of that.

This computes (timestamp - first timestamp) // step for every row instead.

Related issue number

None, found while reading the time series code.

Checks

• I've used pre-commit to lint the changes in this PR (note the same in integrated in our CI checks).
• I've included any doc changes needed for https://microsoft.github.io/FLAML/. See https://microsoft.github.io/FLAML/docs/Contribute#documentation to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

test/automl/test_add_time_idx_col.py fails on main (pandas 2.3 and 3.0) and passes with the change.

[Copilot]

🟢 Approval recommended

The focused fix addresses the reported defects with regression coverage and no identified blocking issues.

No open issues 🎉

What changed in this PR

Fixes time-step indexing for non-monthly/yearly panel forecasting data.

Changes:

• Computes integer time steps for every row, including repeated timestamps.
• Replaces removed pandas Series.view usage with astype.
• Adds regression tests for daily data and shared panel timestamps.

File	Description
test/​automl/​test_add_time_idx_col.py	Tests daily step counts and repeated timestamps.
flaml/​automl/​data.py	Corrects timestamp arithmetic and row alignment.

_{Review effort: Balanced}

[Li Jiang (thinkall)]

Overall review of the complete PR: no blocking code issue found; approval withheld pending CI.

The corrected subtraction/division order computes elapsed time steps, and assigning from the original timestamp column preserves repeated panel dates and row-index alignment. Both focused regression tests passed locally. The review-code harness is clean and Copilot's current Balanced review recommends approval with no open issues. The current head still has failing build checks, so the all-CI-passed approval gate is not met. This review does not approve the PR for merge.

Posted by thinkall-agent-auto-reviewer

[Copilot]

🟢 Approval recommended

The focused fix and regression tests address the reported failures with no identified blocking issues.

No open issues 🎉

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

The focused fix addresses the reported failures with regression coverage and no unresolved findings.

No open issues 🎉

_{Review effort: Balanced}

[Li Jiang (thinkall)]

Overall review at c99d0484e355d7df37ffbfd8fa0060e133ebab21: the ordinary daily/panel correction is useful, but one P2 correctness issue remains in the new step mapping. Complete regular month-end and timezone-aware daily series still receive duplicate indices and artificial gaps; the inline review includes independently executed reproductions.

Validation: python -m pytest test\automl\test_add_time_idx_col.py -q --disable-warnings passed (2 passed). Separate repeated-panel calendar/DST fixtures fail the consecutive-step invariant, and the review-code whole-PR Low harness independently identified the same root cause. Green CI and a positive Copilot recommendation do not resolve this source defect, so approval is withheld. No contributor code was modified.

Posted by thinkall-agent-auto-reviewer

[Copilot]

🟢 Approval recommended

The focused fix and regression tests address the reported failures, with no blocking issues identified; tests were not executed during review.

No open issues 🎉

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Irregular daily data still loses steps across DST, and the month-end test uses an alias unavailable in older supported pandas versions.

1 open issue

• Use compatible MonthEnd offset instead of pandas 2.2-only ME frequency · New

1 previously missed issue

In code that hasn't changed since last review

DST causes irregular daily data to lose calendar-day spacing

flaml/​automl/​data.py:266

The irregular fallback still loses calendar-day steps across DST. For New York midnights on March 12, 13, 14, 15, and 17, 2021, infer_freq returns None. The modal interval is 24 hours, so this division produces [0, 1, 2, 2, 4] instead of [0, 1, 2, 3, 5]. Distinct dates then share a time_idx. Preserve calendar-day spacing for irregular daily data without changing elapsed-time indexing for intraday data, and add a regression combining DST with a missing date.

_{Review effort: Balanced}

[Li Jiang (thinkall)]

Reviewed e796511b5c07e38efe7a190063396202aa98f934 with review-code (Low, gpt-6.1-sol). The complete regular month-end/DST mapping is repaired on modern pandas, but changes are requested for supported-version compatibility.

[P2] Keep the new month-end fixture within the declared pandas range. Both AutoML and test extras permit pandas>=1.1.4, while the new freq="ME" fixture requires pandas 2.2+. I reproduced ValueError: Invalid frequency: ME natively with Python 3.10/pandas 1.5.3. pd.offsets.MonthEnd() produces the same six dates on pandas 1.5.3, 2.3.3 and 3.0.5. This corroborates the existing Copilot finding; no duplicate inline is being added.

The actual focused file reports 5 passed on pandas 3.0.5, but 2 failed, 3 passed on pandas 1.5.3. The second failure is the timezone test: the unchanged pd.infer_freq(Series) call rejects timezone-aware Series on that runtime. Exact prior and current code both show that TypeError, so I am not attributing it to the new ordinal branch. The same dates wrapped in a DatetimeIndex infer D; please ensure the newly added DST regression also runs across the declared pandas range.

Fifteen native calendar/DST/permutation controls verify the modern regular-grid fix and original panel-row/index preservation. Existing irregular modal-gap/DST limitations are recorded separately and are not claimed fixed or introduced by this correction. I have not resolved my older signed calendar/DST thread while supported-version handling remains incomplete, and no other reviewer's thread was touched.

No contributor implementation/test or installed package was changed. Approval remains withheld.

Posted by thinkall-agent-auto-reviewer

[Anish Mehta (anishmehta24)]

Li Jiang (@thinkall) fixed in 0755530. The test uses pd.offsets.MonthEnd(), and infer_freq now gets a DatetimeIndex, so the tz-aware DST case also works on old pandas. All 5 tests pass on pandas 1.5.3 (Python 3.10) and 3.0.6.

[Copilot]

🟢 Approval recommended

The focused implementation and regression tests address the reported failures, with no unresolved findings from static review.

No open issues 🎉

1 resolved since last review

• Use compatible MonthEnd offset instead of pandas 2.2-only ME frequency

_{Review effort: Balanced}

[Li Jiang (thinkall)]

Reviewed the complete PR at 0755530, including the older-pandas correction.

The portable MonthEnd fixture and DatetimeIndex inference fix the two concrete compatibility cases previously reported. The focused file now passes all five tests in each of four existing environments: Python 3.10/pandas 1.5.3, Python 3.10 and 3.12/pandas 2.3.3, and Python 3.13/pandas 3.0.5.

One residual P2 is detailed inline: a complete spring-DST business-day panel still gets artificial step gaps on pandas 1.5.3. Extended native probes passed 59 of 60 regular-calendar evaluations; the failing case makes the existing four-step panel holdout predicate select only three observations per group. This is an incomplete supported-range calendar correction, not a claim that earlier legacy heads handled that input successfully. The already-known irregular modal-gap limitation is not being reposted as a newly introduced defect.

Validation used the real helper and the existing holdout predicate; it did not train TFT or claim a full minimum-dependency matrix. No contributor source or tests were edited.

Posted by thinkall-agent-auto-reviewer

[Copilot]

🔵 Needs a closer look

The irregular-data fallback still produces incorrect time indices for uneven gaps and daylight saving transitions.

No open issues 🎉

2 previously missed issues

In code that hasn't changed since last review

Use a GCD base step to preserve irregular timestamp gaps

flaml/​automl/​data.py:271

For irregular dates such as Jan 1, 3, 5, and 6, frequency inference returns None and the modal gap is two days. The floor division then produces [0, 1, 2, 2], assigning the same time_idx to two distinct timestamps within a series instead of preserving gaps. Derive a base step that divides every observed delta (for example, their greatest common divisor), and add a regression where a shorter gap occurs less often than the modal gap.

Preserve calendar-day indices across DST transitions

flaml/​automl/​data.py:271

For local midnights on March 11, 12, 14, 15, 16, and 17, 2021 in America/New_York, frequency inference fails and the UTC-based fallback produces [0, 1, 3, 3, 4, 5] instead of calendar-day indices [0, 1, 3, 4, 5, 6]. Changing the modal interval to a greatest common divisor prevents collisions but still measures hours rather than calendar days. Preserve calendar-day steps for irregular daily input without changing elapsed-time semantics for subdaily input, and add a missing-date regression across DST.

_{Review effort: Balanced}

[Li Jiang (thinkall)]

Reviewed the complete PR at c88d7e0, including the new wall-clock fallback.

The previous spring-business/DST case is fixed: all six focused repository tests pass in each of four existing environments (pandas 1.5.3, 2.3.3 and 3.0.5), all 60 regular-calendar probe evaluations pass, and the existing period=4 predicate now selects four business observations per group.

One new P2 regression is detailed inline. A genuinely missing elapsed hour during the fall DST fold is preserved by the exact previous helper but compressed away by the current fallback. The paired preservation probe fails on the new source in all four environments. Please keep the business-day correction while validating or narrowing wall-clock inference so irregular subdaily grids retain their gaps.

The independent Low review returned issues_found. Native validation used the real current helper and exact prior-helper controls; no TFT training, public AutoML fitting, dependency changes or full minimum-dependency matrix is claimed. Existing irregular modal-gap/calendar-day limitations were not reposted as newly introduced by this commit.

Posted by thinkall-agent-auto-reviewer

[Copilot]

🟡 Changes recommended

The irregular-series fallback can assign identical indices to distinct timestamps when the most common interval exceeds the shortest interval.

1 open issue

• Use a base step that preserves irregular timestamp gaps · New

_{Review effort: Balanced}

[Li Jiang (thinkall)]

Re-reviewed commits c88d7e0ef1067a05f9bea4306b3155a8f7a43d96 -> 4dc11355d4b0f52660098c01fab79c53bb0605ab. The only delta is main's website/yarn.lock update of follow-redirects 1.16.0 -> 1.16.1. One explicitly head-bound review-code Low execution (actual runtime gpt-6.1-sol) found no new issue in this identical inherited patch. Controller Git comparisons verify all non-lockfile tree entries are unchanged. This is a scoped merge-delta review, not a fresh clean signoff for unchanged feature code.

Version/URL/caret/SRI shape checks pass. No independent local registry/tarball verification, package smoke test or website rebuild is claimed. Previous feature findings and runtime validation limits remain preserved.

Changes remain required. The existing P2 fall-DST regression remains: removing the actual repeated 01:00-05:00 hour from the New York hourly grid should preserve [0, 1, 3, 4, 5, 6], but the wall-clock fallback compresses it to [0, 1, 2, 3, 4, 5]. That exact prior/current paired failure was reproduced in all four existing environments before this merge. The helper is byte-unchanged, so this is retained evidence, not a claim of four new runs. The original finding is not resolved by this dependency update.

At the 2026-10-26T12:13:44.943946+00:00 gate snapshot, actual current-head workflows were: docs completed/success; Code formatting completed/success; Build in_progress/no conclusion. Copilot has Balanced review without an Approval recommended verdict; 2 review thread(s) remain unresolved. Gate approval is not successful CI completion, and no new PR approval is being submitted while required gates remain unmet.

Posted by thinkall-agent-auto-reviewer

[Copilot]

🟢 Approval recommended

No active review issues remain for the time-index generation changes.

No open issues 🎉

1 resolved since last review

• Use a base step that preserves irregular timestamp gaps

_{Review effort: Balanced}

[Li Jiang (thinkall)]

Reviewed the complete current PR at 597e23191befaa789b63b80faca940bf3adcc9e7, including the latest calendar/GCD correction.

The previous missing-fold-hour finding is fixed: the new source preserves [0,1,3,4,5,6], and the exact previous helper fails that regression. I verified the correction on both native stacks and resolved only my corresponding owned thread.

Two P2 issues remain, detailed inline: deleting one month-end observation changes month-relative steps into day-relative steps, and the wall-clock grid validator can allocate millions of candidate timestamps for only three observations.

Validation: all nine repository helper cases pass on each of pandas 2.3.3/NumPy 1.26.4 and pandas 3.0.5/NumPy 2.3.3 (18 executions, nine distinct cases). Another 73 repeated/shuffled panel fixtures per stack and nine unit/GCD/duplicate-index controls per stack pass. The separate exact prior/current missing-month counterexample and real bounded allocation witness establish the two findings despite that selected suite passing.

Whole-feature review-code effort was Medium: six qualified independent grok-4.7/xhigh lenses. Seven actual executions and all original results are retained in locally committed canonical evidence; one initial lens was excluded after its grep exposed sibling feedback and was independently rerun in a fresh source-only checkout. Neither excluded feedback nor a scoped correction is presented as independent clean signoff.

This is helper/panel-index validation, not estimator fitting or TFT training. No full minimum-dependency matrix, native Spark, pandas 3.1 execution, giant nanosecond-grid allocation or actual OOM is claimed. Historical version-compatibility and short-series allegations were adjudicated against concrete controls rather than automatically reposted.

At 2026-10-30T16:41:50.130084+00:00, actual current-head pull_request workflows were: docs completed/success; Code formatting completed/success; Build completed/success. Current Balanced Copilot review 5714447225 recommends approval. These external gates do not clear the two independently verified source findings, so changes remain required.

Posted by thinkall-agent-auto-reviewer

[Copilot]

🟡 Changes recommended

The helper assigns distinct daylight-saving fold observations the same time_idx because it does not require strictly increasing month ordinals before bypassing the elapsed-time fallback.

1 open issue

• Reject duplicate month ordinals during frequency inference · New

_{Review effort: Balanced}

[Li Jiang (thinkall)]

Re-reviewed the correction from 597e23191befaa789b63b80faca940bf3adcc9e7 to f4c65a4349354907269129f339e4e8039018af01, and composed it with the retained whole-feature review rather than treating a scoped correction as blanket clearance.

Both previous P2s are fixed. Sparse month/quarter cases now retain calendar steps, and the grid validator builds at most one candidate per observation. A real nanosecond fold-gap probe now creates exactly three entries (24 bytes), with no giant allocation. Only my two corresponding signed threads have been resolved after committed native fix proof.

Changes remain required for the existing Copilot finding; I am corroborating it instead of posting another inline. New York instants 2021-10-07T01:30-04:00, 2021-11-07T01:30-04:00, and 2021-11-07T01:30-05:00 are distinct. The exact previous helper returns [0,744,745] per panel group, but the new calendar candidate returns [0,1,1], merging the two fold observations. I reproduced this with the real current helper and exact prior controls on both native stacks. Reject calendar candidates with non-increasing month ordinals so these inputs reach the elapsed-time fallback, and add the repeated/shuffled-panel regression.

Validation: all 12 repository cases pass on each of pandas 2.3.3/NumPy 1.26.4 and pandas 3.0.5/NumPy 2.3.3 (24 executions, 12 distinct cases); another 97 panel/calendar/unit fixtures pass per stack (194 evaluations). Exact prior controls give nine passes and three expected failures per stack, proving the targeted corrections. The separate fold counterexample still fails the timestamp-to-index uniqueness invariant on current source.

review-code Low used the actual CLI-default gpt-6.1-sol and returned issues_found. Its canonical evidence is committed locally on this exact author source. The previous whole-feature Medium's six qualified independent lenses, excluded initial result and validation limits are retained; this is not six new votes or a new compatibility guarantee. No estimator fitting, TFT training, native Spark, pandas 3.1 execution, full minimum-dependency matrix, huge allocation or OOM is claimed.

At 2026-10-30T21:15:36.291172+00:00, actual current-head pull_request workflows were: Build completed/success; docs completed/success; Code formatting completed/success. Current Balanced Copilot review 5718569371 says Changes recommended. The verified collision and its existing Copilot thread remain blockers; no PR approval is being submitted.

Posted by thinkall-agent-auto-reviewer

[Copilot]

🟢 Approval recommended

There are no active review issues for the time-series index generation fix.

No open issues 🎉

1 resolved since last review

• Reject duplicate month ordinals during frequency inference

_{Review effort: Balanced}

[Li Jiang (thinkall)]

Completed an overall review of the current PR at 0fe1cfdfb33c50056e6150e96af94130ea1b7a65, composing the complete current source with the retained whole-feature review and the new corrective evidence. No actionable issue remains in that qualified assessment.

The duplicate-month/fold finding is fixed. Both exact prior failing panels now preserve distinct timestamps: [0,1,1] becomes [0,744,745], and [0,0,1,2] becomes [0,1,721,1465]. Repeated and shuffled panel rows are covered. The helper rejects non-increasing month ordinals and reaches the existing elapsed-time fallback rather than merging fold observations. Sparse month/quarter and the earlier daily/business-day/DST and bounded-grid corrections remain intact.

Native validation on Windows pandas 2.3.3/NumPy 1.26.4 and pandas 3.0.5/NumPy 2.3.3: all 14 repository cases pass per stack (28 executions, 14 distinct cases), plus 97 regular/irregular/calendar/unit fixtures per stack (194 evaluations). Identical final repository cases with the exact previous helper give 12 passes and two expected failures per stack. The real nanosecond grid probe still creates three entries (24 bytes); no huge allocation or OOM was attempted.

The actual CLI-default gpt-6.1-sol Low corrective review returned clean; its canonical evidence is committed locally on this exact source parent. This is not six renewed whole-feature votes: the previous six qualified independent Medium lenses, excluded initial result, later corrective history and driver whole-source composition are retained with their original outcomes. The prior Copilot thread was resolved externally, not by this automation.

This is qualified native helper/panel-index validation, not a claim of public AutoML fitting, TFT training, native Spark, pandas 3.1 execution, the full minimum-dependency matrix, or universal compatibility beyond the inspected paths.

Fresh same-head gate check at 2026-10-31T01:50:45.317653+00:00: docs completed/success (attempt 1); Code formatting completed/success (attempt 1); Build completed/success (attempt 1). All review threads are resolved, current Balanced Copilot review 5720361750 explicitly recommends approval, all applicable contexts are successful, and all actual applicable workflow runs completed successfully. The gh-release job is verified PR-inapplicable at the unchanged workflow blob; its skipped result is not counted as passed validation. Approving the complete PR with the validation qualifications above.

Posted by thinkall-agent-auto-reviewer