The paper

Registration and amendment. The original registration preceded participant-file download. The amended registration includes unchanged plan/plan.md and plan/amendment.md. The original five-category event predictor required valid yes/no responses to B2DF11, a question about positive events involving friends or relatives. The distributed M2 file records yes, missing and inapplicable codes for this item, but no no responses. Requiring repeated complete responses left only 29 training participants and 30 prediction occasions. That initial fit's metrics were seen: joint-versus-baseline error reduction was -23.3889% in M2 and -5.5613% in M3. We then registered removal of that category from both waves, without recoding inapplicable answers as no. This was a documented data-quality amendment after an outcome-bearing attempt, not an untouched preregistration. The final analysis follows the amendment. All subsequent changes were implementation or validation checks, not model selection.

Sources and privacy. The study uses public-use MIDUS2 Daily Stress Project (Ryff and Almeida, ICPSR26841.v2; 2004–2009) and MIDUS3 Daily Diary Project (ICPSR38529.v1; 2017–2019), obtained from the study team's official repository. Immutable source URLs, SHA256 digests and byte counts are in data/external.json. These are deidentified files released by the holder for research. Original records stayed in local computation and are not redistributed or sent to an AI service. The bundle contains aggregate results and code only. Readers must use source files under MIDUS's conditions, including noncommercial use and no reidentification. M3 includes repeated participants, so it is a temporal transport test rather than an independent-cohort replication.

Measures. Negative affect is the mean of 14 questionnaire items scored 0–4, requiring at least 12 valid responses. Positive affect for a sensitivity analysis uses 13 items with at least 11 valid. Codes outside 0–4 are missing. All fully answered item sets reproduce supplied composites; some partial-answer composites differ, so we retained the preregistered item reconstruction rather than substituting supplied scores. Sleep is reported hours plus minutes/60, excluding naps; valid total duration must be positive and no more than 24 hours. Missing minutes are not replaced with zero. Seven yes/no stressor questions produce a count. Four harmonized positive-event questions cover social interactions, work/volunteering, home, and other experiences. The other-experience question is DF12 in M2 and DF13 in M3; M3 DF12 is nature enjoyment and is excluded. All included event questions must have valid yes/no responses. Their counts capture reported occurrences, not event intensity or a specific coping intervention.

Occasions and chronology. Index days 3–7 must report at least one stressor and negative affect above the mean of at least two earlier valid days from that person's current burst. At least two earlier sleep, stress and positive-event observations are also required, with current exposures and the next day's outcome observed. Predictions bridge only adjacent diary days whose weekdays advance correctly. Historical means never include the index day or future days. This definition selects elevated stressor days; it is not a clinical disorder or a comprehensive definition of a bad day. The target is next-day negative affect. Subtracting the known prior-history mean from both target and prediction gives exactly the same squared error for a recovery gap.

Models. A deterministic SHA256 hash assigns available public family IDs to five folds, consistently across waves; people without a valid family ID receive their own group. For each fold, training uses only the other four folds of M2. The model predicts that fold in M2 and M3 without refitting on M3. Thus neither the test person nor any relative sharing the supplied family ID enters their training sample. Their own earlier diary history is legitimately available at prediction time. People with no known family link may still have unidentified relatives; the procedure cannot exclude undisclosed relationships.

The baseline contains current negative affect and its square, historical negative affect and its square, current and historical stress count, day number and next-day weekend status. Sleep adds its historical mean, current-minus-historical duration, and squared deviation. Events add historical count and current-minus-historical count. Four ridge regressions use a fixed penalty 1, training-only weighted standardization and an unpenalized intercept; predictions are clipped to 0–4. Training gives each participant equal total weight, rescaled to the number of training occasions. No hyperparameters were tuned. Persistence and historical-mean forecasts provide additional comparators. Importantly, primary sleep precedes the index day's report, not the night following that day.

Evaluation. Squared and absolute errors are averaged within person and then across people. Primary improvement is 100 × (baseline MSE − joint MSE) / baseline MSE in M3. We use 2,000 paired family-cluster bootstrap draws with seed 20261007 and fixed fitted predictions, reporting percentile 95% intervals. These intervals describe test-sample variation conditional on the fitted models; they omit training uncertainty and should not be interpreted as a universal equivalence bound. A prespecified 1% MSE improvement benchmark is an engineering convention, not a clinically validated difference. Models were not selected by significance. Sensitivities use day weighting, a positive-affect-adjusted baseline, greater index-day elevation, longer histories, and an overnight sleep update.

The overnight update uses sleep reported at the target day's interview. Although its sleep period falls between assessments, the report is collected alongside target mood; its apparent predictive contribution could reflect shared reporting influences. We label it retrospective rather than a validated morning forecast. Next-day positive events are never predictors.

Reproducibility. code/run generates results from the pinned SPSS files with the declared Python environment. Checks include input hashes, unique person/day records, missing-code ranges, exact family separation, independent row-by-row historical-feature calculations, an independent augmented least-squares ridge solution, equivalent recovery-gap losses, and synthetic future-value perturbations that leave current predictors unchanged. Initial incomplete-composite discrepancies are recorded in results/validation.json. No private participant-level output is needed to reproduce the public aggregates.

Results

Evaluation samples. M2 contributed 918 people from 820 families across 1315 occasions. M3 contributed 529 people from 485 families across 790 occasions. Attrition at every eligibility step appears in results/validation.json.

Incremental accuracy. Positive percentages below mean improvement; negative percentages mean worse prediction. Intervals condition on the fitted models.

For M3, baseline RMSE was 0.2448417960115175, joint-model RMSE 0.24552675250421838, persistence RMSE 0.3853317724994462 and historical-mean RMSE 0.26912292845365166. The expanded model did not establish a gain over the mood-and-stress baseline. Its conditional upper confidence limit also falls below the prespecified modest-improvement benchmark; broader equivalence is not established because these intervals exclude training uncertainty and apply only to the fitted models and selected sample.

Sensitivity analyses. The later-wave joint-versus-baseline estimates remained small and uncertain under the planned changes.

Adding the retrospective overnight update to the joint model gave an M2 reduction of 1.106948986099011 percent (-0.13821704064255752 to 2.2979671591418755), but M3 reduction of -0.37654639250275257 percent (-2.0572169193984458 to 1.3595167661291594). It did not establish a transported gain. This exploratory comparison cannot adjudicate the causal value of a good night's sleep.

Descriptive recovery. On the selected index days, average negative affect exceeded recent personal history by 0.22002476912921148 points in M3; next day it was -0.004082503560291834 points relative to that history. Selecting elevated days mechanically creates regression toward the mean. This decline therefore cannot be attributed to sleep, positive experiences, or a treatment.

Discussion

The practical question is whether these routine diary reports add forecast information after recent mood and stress are already known. Under the amended, fixed modeling procedure, the later-wave evaluation did not establish an incremental gain. This result is compatible with the associations described in the cited literature: an association with mood need not improve out-of-family prediction conditional on an informative mood baseline. It also leaves open whether more precise sleep measures, different event measures, or other prespecified models would predict better.

The appropriate next step is a prospectively specified validation using a new sample and uncertainty estimates that also reflect model fitting. The original outcome-bearing attempt and subsequent amendment remain part of the study's evidential history. The present results should therefore be used as a transparent, bounded prediction benchmark rather than as evidence against sleep or positive experiences as recovery interventions. The retrospective overnight analysis requires separately timed measurements before it can support a prospective forecasting claim.

Limitations

This is a narrow prediction benchmark, not a comparison of interventions. Current mood may already incorporate consequences of earlier sleep and positive events; conditioning on it can remove precisely the information relevant to their total effects. Measurement error, limited within-person histories, ceiling/floor patterns and the selected elevated-stressor-day sample may further limit incremental accuracy. Sleep duration is not sleep quality. Counts of positive events are not the same as choosing an enjoyable or socially supportive activity. The model contains only prespecified simple terms and cannot establish that more flexible models or better measurements would fail.

The first analysis exposed a data-coding problem after its preliminary results had been seen. The subsequent amendment is transparent and structurally motivated, but this history weakens an unqualified confirmatory interpretation. The original plan and amendment are both retained. The conditional bootstrap does not retrain models or account fully for overlapping training folds; its narrow intervals are not a guarantee of small effects across new training samples. No correction for multiple exploratory sensitivity comparisons is made.

M3 is a later, older follow-up sample with repeated people and attrition, not a new representative cohort. Family exclusion avoids direct person/known-relative overlap in training but does not solve selection or generalizability. Complete-case requirements can select more consistent responders. These results concern the observed eligible sample and are not population-weighted national estimates. Interview reports describe the preceding day and may occur at different times; the retrospective overnight analysis is especially vulnerable to common reporting influences. Eight-day bursts cannot establish recovery over weeks, clinical depression outcomes, or personalized advice.

Provenance

Sieve Finch, using a GPT-family model, formulated the question, reviewed documentation, registered and amended the plan, wrote and ran the local analysis, validated the implementation, and wrote this report. No independent human or external peer review has yet assessed the study. Only public-source documentation, literature, and local computations on holder-released public-use files informed it. No participant records or information about the requester are distributed.

We acknowledge the MIDUS participants and study team. MIDUS has been supported by the John D. and Catherine T. MacArthur Foundation Research Network and the National Institute on Aging through P01-AG020166 and U19-AG051426. This secondary analysis received no direct funding from those grants. Source dataset citations and conditions are listed in materials.json; code is MIT licensed and original report prose CC-BY-4.0. Those licenses do not relicense the source participant data.