MODELSTIMING_ENGINESMarriage algorithm #3124

STATUS: STABLE

Marriage Timing Engine Results

v3.1.24

Internal diagnostics dashboard for model calibration, error profile review, and intervention planning before rollout.

Runtime Metadata

Last Run

2026-03-15 22:14 UTC

Runtime

00:04:12

Cluster

cf-us-east-2

Window

Trailing 90 days

Cumulative Window Accuracy

Percent of predictions falling within each month-error tolerance.

Within 3M Error

47.3%

5,901 / 12,480 records

+1.2% vs previous run

Within 6M Error

68.9%

8,598 / 12,480 records

+0.8% vs previous run

Within 9M Error

81.7%

10,195 / 12,480 records

+0.4% vs previous run

Within 12M Error

89.8%

11,207 / 12,480 records

+0.3% vs previous run

Absolute Error Distribution

Spread diagnostics for monthly prediction error.

Min / Max Error Split

Minimum Absolute Error

0.14 months

Case STB-4044

Maximum Absolute Error

11.62 months

Case OTL-8872

Median Absolute Error

2.06 months

Median over entire validation slice

Std Deviation

2.94 months

Broadness of model spread

Dataset Statistics

Current evaluation slice and distribution overview.

Evaluated Rows

12,480

post-sanitization pairs

Unique Couple IDs

8,912

de-duplicated entities

Training Horizon

2009-01 to 2025-12

monthly windows

Outlier Share

10.2%

absolute error > 9 months

Suggested Algorithmic Improvements

Manual dev notes from the latest error audit.

Suggested manual intervention

Early marriage cohorts (age bucket 20-24) remain slightly over-predicted in urban clusters. Review feature weighting for relocation recency and second-house dasha interactions before next calibration.

Agentic Rule Proposal

Heuristic generated from internal simulation replay.

Agentic optimization suggestion

Expected Lift +2.3% to +3.4%

Apply a dynamic correction factor when the model confidence is low and relocation signal spikes in the last 18 months.

IF abs(relocation_volatility_18m) > 0.62
AND confidence_score < 0.71
AND age_bucket IN ["20-24", "25-29"]
THEN predicted_marriage_months = predicted_marriage_months - 1.1
ELSE keep_original_prediction

Promote to experiment bucket only after shadow validation on the next 2,000 rows.

Case-Level Diagnostics

Outliers and exact-match cohorts from this run.

Erroneous / Outlier Cases

5 rows after search filter

Case	Cohort	Pred	Actual	Abs Err	Conf	Note
OTL-8872	US-West / 25-29	2027-04	2026-04	11.62M	0.59	Late relocation signal underweighted
OTL-7741	US-East / 20-24	2026-11	2026-01	10.07M	0.56	High volatility in career transition
OTL-6538	IN-Metro / 30-34	2025-09	2026-06	9.78M	0.61	Transit shock not represented in base rule
OTL-5299	US-South / 25-29	2028-02	2027-05	9.34M	0.54	Sparse historical analogs for profile
OTL-4920	IN-Tier2 / 20-24	2026-12	2026-03	9.12M	0.58	Weights overfit to delayed cohorts

Successful / Exact-Match Cases

5 rows after search filter

Case	Cohort	Pred	Actual	Abs Err	Conf	Note
STB-4044	US-West / 25-29	2026-09	2026-09	0.14M	0.91	Exact-month hit
STB-3991	IN-Metro / 30-34	2026-05	2026-05	0.21M	0.89	High feature alignment
STB-3810	US-East / 25-29	2026-08	2026-08	0.27M	0.87	Stable dasha + transit window
STB-3774	IN-Metro / 20-24	2026-02	2026-02	0.33M	0.86	Strong historical analog coverage
STB-3601	US-North / 30-34	2026-12	2026-12	0.39M	0.88	Trend-consistent signal profile