XBRL goodwill regressions¶
This exploratory R analysis relates the difference between a target's counterfactual MQ with its acquirer and its observed pre-deal MQ to the goodwill-to-consideration ratio. One command produces the existing full-training results plus separate time-profile in-sample, time-profile out-of-sample, and fixed-2016-MQ results. Each sample writes its own construction fragment and two regression tables.
Inputs
| Dataset | Location | Purpose |
|---|---|---|
| Full-training XBRL deals and counterfactual MQ | /data/<coverage>/analyses/full_training/ |
Existing in-sample comparison; joined by xbrl_deal_target_id |
| Time-profile XBRL deals and counterfactual MQ | /data/<coverage>/analyses/profile_time/training/ |
Profile rows carry sample_status; joined by (xbrl_deal_target_id, sample_status) |
Code and outputs
| Item | Location |
|---|---|
| R module | analyses/modules/xbrl_deals.R |
| Rscript entry point | analyses/pipelines/xbrl_deals.R |
| Slurm task | xbrl-deals |
| Output paths | xbrl_deals section of analyses/config/config_full.yaml |
| Full training | /artifacts/<coverage>/analyses/xbrl_deals/ |
| Time-profile in-sample | /artifacts/<coverage>/analyses/xbrl_deals/time_in_sample/ |
| Time-profile out-of-sample | /artifacts/<coverage>/analyses/xbrl_deals/time_out_of_sample/ |
| Fixed 2016 MQ | /artifacts/<coverage>/analyses/xbrl_deals/fixed_2016/ |
Sample and specifications¶
The sample-construction fragment reports sequential counts for positive
goodwill, both parties mapped to Revelio, available controls, positive
consideration, and at least 20 target MQ score rows. target_mq_n counts
finite relationship-period model scores in the target RCID-year; it is
distinct from target_n_relationships, which counts unique employment
relationships. The main regression sample requires positive goodwill, both
mappings, positive consideration, an observed counterfactual MQ, at least one
observed target relationship, and complete values for the modeled controls.
The second regression table additionally requires target_mq_n >= 20.
The dependent variable is the winsorized log of goodwill divided by
consideration transferred. The main regressor, Delta MQ, is counterfactual
target MQ minus observed target MQ; pre-deal acquirer MQ is also included in the
controlled specifications. Full-training and time-profile tables use five
specifications: Delta MQ alone, core controls, all controls, year fixed
effects, and year fixed effects plus separate acquirer and target industry
fixed effects. Industry fixed effects use the first digit of the NAICS prefix
stored as industry2. The fixed-2016-MQ/2017-deal table uses the same first
three specifications and a final specification with acquirer and target
industry fixed effects, but no year fixed effect because every deal is from
2017.
Standard errors are clustered by acquirer CIK. The year fixed effect uses the
selected goodwill fact's reference year. The regression year range is therefore
not the same as the MQ year range.
The time-profile in-sample tables use the in_sample deal/status row across
the available in-sample MQ years. The out-of-sample tables use the
out_of_sample row, where MQ is measured through 2024. The fixed-2016 test
selects the in-sample row with observed_mq_year == 2016 and restricts the
goodwill-event year to 2017. It therefore uses 2016 worker embeddings for the
counterfactual; the acquirer company embedding is fixed. Compustat controls
remain the latest annual values available by the deal date, consistent with
the current deal panel.
MQ year coverage
All contexts include observations through observed_mq_year == 2024.
This ceiling applies to the pre-deal MQ year, not the goodwill-event year.
Thus 2024 deal events using 2023 MQ and 2025 deal events using 2024 MQ are
eligible.
Running the analysis¶
Build the base deal regression data and counterfactual MQ file first. Then run the regressions from the analyses directory in the analyses container:
The Slurm task requests 8 CPUs, 64 GB RAM, and 4 hours by default; standard
resource overrides such as --time, --mem, --cpus, --partition, and
--dry-run are available through slurm/submit.sh. Each available context
rewrites its own configured LaTeX fragments and refits the regressions. If
the full-training input files are unavailable, that context is skipped and
its existing result files are left untouched. The analysis does not cache
fitted models. This remains an exploratory analysis pipeline.