Beyond Prediction: Prescriptive AI for Strategic Energy Capital Allocation

A governed decision intelligence system for risk, returns and investment-committee choices across oil, gas, power, renewables and storage

SIADS 699 Capstone — Master of Applied Data Science, University of Michigan — Group 13 (solo project). Author: Alan Mössinger. Mentor: Laura Danielle Stagnaro. November 2026.

The deliverables

Deliverable	Where
The report	report/Beyond_Prediction_Edited_Full_Report_V2_revised.pdf (the Word source: Beyond_Prediction_Edited_Full_Report_V2_revised.docx) — abstract, an executive summary table, twelve sections within the course's 3,000 words, statement of work, appendices A–M, references. Every number in it is checked against results/report_numbers.csv by tests/test_report.py, and so is the body's length
The poster (A1, one page)	report/poster_V2.pdf, drawn from the report's numbers by report/poster_V2.typ (make poster); every number on it is checked by the same test
The code	thirteen notebooks run in order by make all; the src/ package; the test suite (make test)
The dashboard	make app — three pages: Decide, A worked example, The report's decision (see below)
The proposal and the evaluation plan	docs/SIADS699_Project_Proposal_Group13_2026-10-19.pdf, docs/proposal_update_2026-10-19.md; docs/SIADS699_Evaluation_Plan_Group13.pdf, docs/evaluation_plan.md
Mentor feedback and my reply	docs/mentor_feedback.md (every item, what changed, where), docs/mentor_reply_2026-10-27.md
The record of the work	docs/decisions.md (127 numbered decisions with reasons), docs/registration_ledger.md (every pre-registered line, its verdict as first scored, every correction); the index is docs/README.md

The question

Energy companies allocate one capital budget across investments — oil and gas, conventional generation, renewables and storage — that differ in return, market exposure, execution risk and downside. Forecasts and project valuations rank these investments one at a time; they do not show that a set of individually attractive projects can carry a concentrated tail, an unplanned call on liquidity and an exposure to one adverse market. The question this project answers is prescriptive: given limited capital, uncertain markets, competing investments and stated risk preferences, which projects should a company fund, stage, defer or reject — and why?

The system runs in six stages, kept apart so that each one's contribution can be measured on its own: prediction (which interconnection requests get built and when; lease output, plant output and prices), uncertainty (2,000 joint scenarios of prices, output, costs and timing), prescription (a risk-aware optimizer choosing a book under a budget and caps), AI challenge (a language model that reads the committee's reasons into rules and argues against the proposal from computed facts, holding no authority), human decision (an override gate and a hash-chained ledger) and measured outcome (pre-registered criteria, scored once).

The decision, and the answer

The case: on 1 January 2025 a company decides one budget of $3.82B across twenty producing Gulf of Mexico leases to redevelop and twenty ERCOT interconnection requests, built only from what was public that day, scored on 2,000 held-out scenarios against criteria registered before scoring.

• Fund the risk-aware book — five leases and eight requests, $3.78B committed ($1.42B of lease capital, $7M of deposits, a $2.35B construction reserve) — together with a $2.53B contingent construction facility, because construction calls exceed the reserve in 50% of scenarios. Defer the 12 waiting requests until their agreements are signed; monitor 14 leases; reject 1.
• On the registered risk-adjusted metric (expected NPV minus the average loss in the worst 5% of scenarios) the risk-aware book beats ranking by NPV by +$219M [+$13M, +$401M] as pre-registered, and by +$312M [+$110M, +$506M] once every decision process is financed the same way (+$213M [+$17M, +$412M] over the deterministic optimizer). It expects $2.17B with a $0.82B worst-5% average loss and a 9% chance of a loss ($1.98B and $0.95B financed), and it scores best in 7 of 7 modelled worlds. Without a facility, equal capital by sector would lead: the recommendation is the book and its facility.
• Better prediction did not make a better decision. At utility scale the completion model ties a size lookup at entry; the book chosen with the model's queue curves is not demonstrably better than the book chosen with the lookup's. The decision turns on a few large requests, on the capital plan and on the tail, not on the forecast.
• The AI's value is in the argument, not the choice. In a registered test on 48 simulated committee proposals, deterministic tools improve the committee's own decision by +$76M per case and the language-model challenger adds +$43M [+$9M, +$86M] more; against a fixed optimizer-first rule that reads no words the model's edge disappears (−$14M [−$36M, +$2M]). The model named 12 of 12 author-written flaws of argument and over-challenged 0 of 12 sound proposals; every number it writes is checked against its fact sheet by code.
• Reported as they fell. 41 criteria and 87 lines were registered; 40 lines pass. No registered verdict was re-scored; corrections are added beside the original in the ledger. Of the 12 controls of the governance layer, 10 are in place and 2 partly (injection resistance, scenario realism). All results are model-based on public data and assumed lease costs, not realized returns.

How the work is organized

Each notebook has the same shape: a title cell that says its step in the pipeline, what it reads, what it writes and which notebooks read it; an inputs-and-outputs table; numbered sections; a Reading this cell under every table and figure; and takeaways that say what goes to the report. The larger notebooks are split into lettered parts (part A, part B, …) whose sections are numbered A.1, B.2, …, and cross-references use that form (notebook 04 §J.3 is section 3 of part J of notebook 04). make all runs the notebooks in dependency order (01, 03, 02, 04, 05, 07, 08, 09, 06, 10, 11, 12, 13) and writes every number the report cites to results/. The report's body rests on notebooks 11–13, which read everything upstream; its appendices use figures from every notebook.

Notebook	What it does	What the report takes from it
01_problem_and_data	the queue sources (LBNL Queued Up 2025 and 2026, EIA-860M, the ISO files), the energy sources (BSEE, EIA-923, prices, ERCOT, eGRID), the cleaning rules with row counts, the manifest, the data-readiness gate	Section 3 (data), Appendix E
02_eda	survival curves, what drives withdrawal, the project map, state archetypes; the energy series	Appendix D (A1, A3, A8, A9), Appendix E
03_statistics_and_features	queue outcomes and entry-time features; leakage-safe features at rolling origins (the truncation proof); criterion C20 registered	Section 3, Appendix E
04_forecast_queue	the completion-and-timing engine: discrete-time competing-risks models against the fair baseline and the size lookup, the level repair, the landmark models, MISO/SPP on the 2026 file	Section 3 (prediction), Appendix E
05_forecast_energy	lease output, plant capacity factors and three prices at rolling origins against five transparent baselines; C20 scored	Section 3 (the forecast contests), Appendix F
06_evaluation_record	every registered criterion scored as registered (C1–C24), post-hoc and tournament, explanations, the decision board; runs after 09	Section 8, Appendix A, Appendix D
07_economics	the forty candidates' cash flows and NPV under the documented cost grid, with a provenance label on every value	Section 2, Figure K1
08_scenarios_and_risk	joint price, output, cost and timing scenarios with mean reversion; named stresses; risk measures per candidate	Section 4, Figures A5, A6, K2
09_portfolio	the four books under one budget (NPV ranking, equal capital, the deterministic and the risk-aware optimizer); C21–C24	Section 5, Appendix E
10_advisor_and_gym	the grounded language-model advisor, the challenge round and the override gate, the decision gym on sixteen past situations, the forward view at 1–20 years	Section 7 (the gym), Appendix F
11_decision_system	the frontier over tail weights, what-ifs, value of information, staged decisions, the processes compared on held-out scenarios, the active challenger; the as-of candidates	Figures 1 and 4, Appendix B, Appendix D
12_three_questions	the registered experiment on 48 simulated committees: the four-arm ladder (proposal → tools → tools + model → optimizer), paraphrases and injections, the credit line priced, the one-year forward test; C33–C41	Figure 3, Tables 2 and 3, Appendix D
13_executive_corrections	every process financed the same way, the co-movement dial and seven worlds, the optimizer-first rule as a fifth arm, the allocation line by line, the committee workflow and the control matrix, an emissions cap	Figures 2, 5 and 6, Table 1, Figures K3–K5

src/ holds the functions the notebooks share, each module with a docstring that says what it does, why, what it reads and writes, and who uses it:

Group	Modules
Data and cleaning	get_lbnl, get_data, eia_api, ercot_api, ercot_history, energy_data, iso_files, audit_queue, cleaning, linkage, geo, asof_candidates, decision_date
Queue models and their evaluation	events, features, discrete_time, survival_tools, baselines, benchmarks, drift, drivers, landmark, variants, track_b, explain, evaluation, posthoc, forward
Energy forecasts	decline, energy_features, energy_models, energy_stats
Economics, scenarios and the book	value, economics, valuation, scenarios, optimizer, portfolio, robustness, whatif, voi, dynamic, horizons
The AI challenger and governance	advisor, advisor_tools, challenge, challenge_experiments, challenge_v10, gym, ledger, registration_ledger, worked_example
Plumbing	paths, config, plots, report_numbers

The dashboard (make app, app/) reads results/ only and recomputes no model. Decide: the options, your book and your reasons, the advisor's challenge with its five reasoning questions, your result against the advisor's solution. A worked example: one executive's recorded round through those steps, for a new reader. The report's decision: the recommendation on one screen in seven tabs, the last one the committee's own call. Challenging your own book needs a language-model key from any provider (see How to run); the recorded rounds replay without one, and the dashboard says so at the top.

Tests (tests/, pytest): the audit, the cleaning rules, the features and the truncation proof, the metrics and baselines, the economics identities, the scenarios, the optimizer's constraints, the advisor's grounding and replay, the gym, the ledger's gate, the challenger, the dashboard pages headless, and that every number in the report and on the poster is in results/report_numbers.csv.

Data

Source	What	Access
LBNL Queued Up: 2025 Edition project file (frozen)	36,441 interconnection requests through end-2024 — the basis of every registered result	download by hand from https://emp.lbl.gov/queues; not redistributed
LBNL Queued Up: 2026 Edition project file (alongside)	38,201 requests through end-2025: reconciliation, the forward test, MISO/SPP dates	same page; CC BY 4.0
EIA-860M via the EIA API v2	inventory of operable U.S. generators (operating month, location)	free API key in .env
CAISO public queue report (via gridstatus)	cluster, study-phase and agreement status of CAISO requests	data/raw/iso/
PJM queue export and ERCOT GIS reports	PJM request, exit and status fields; ERCOT phase dates and the 2024 reports the candidates are built from	fetched from pjm.com and ercot.com; data/raw/iso/, not redistributed
BSEE OGOR-A, EIA-923, EIA and ERCOT prices, NREL ATB 2024, EPA eGRID	lease production, plant generation, the market months, the cost and emissions layer	data/raw/; every file listed with its hash in data/raw_manifest.csv
EIA wholesale prices, PJM market-monitor report, LBNL PJM cost data, IRS Notice 2025-42, Census boundaries	the value and cost layer and the map	data/raw/value/, data/raw/lbnl_costs/, data/raw/geo/

The full data access statement, terms and data dictionary are in data/README.md.

How to run

Requires Python 3.11, GNU make and poppler's pdftotext on the PATH (notebook 07 reads two tables of the PJM market monitor's report from its PDF; pdftotext -v must report a poppler version — the pdftotext that Git for Windows ships is xpdf's and lays the table out differently, so put poppler's directory first on the PATH). On Linux LightGBM also needs the OpenMP runtime (libgomp1).

make setup     # .venv, pinned packages, Jupyter kernel "queue-survival"

1. Get the raw files. data/raw/ is not in the repository (the LBNL file carries no redistribution grant); data/raw_manifest.csv lists all 111 files it must hold, each with its source and SHA-256. With the author's bundle (raw_data_bundle.zip, 137 MB, available on request), unzip it into the repository root and run make check-data, which must end with "111 of 111 raw files present with the recorded hash". Without the bundle, make get-data downloads every file a script can fetch and prints, for each file it cannot, the route in data/README.md. A file whose hash differs from the manifest is reported, never used silently.
2. Copy .env.example to .env. The EIA key is needed only to pull a missing EIA file again. For a live call to the advisor add a language-model key from any provider — ANTHROPIC_API_KEY, or LLM_API_KEY with LLM_BASE_URL and LLM_MODEL for any OpenAI-compatible service — with none, the advisor and the challenger replay their records (results/advisor_runs.jsonl, results/v10/), so make all runs offline. The replay is keyed on the exact prompt text; on another platform, floating-point differences in the last decimals of a table can miss the record for a few calls (decision #111). A clean Linux run from the bundle reproduced 2,548 of the 2,561 numbers the report cites exactly as printed and every registered verdict (decision #112).
3. Run:

make all       # notebooks 01 → 13 in dependency order, stopping at the first error (about three hours)
make numbers   # results/report_numbers.csv, every number the report cites
make test      # pytest
make poster    # report/poster_V2.pdf from report/poster_V2.typ (needs typst)
make app       # the dashboard at http://localhost:8501 (make dashboard is the same; on Windows, double-click dashboard.bat)

In VS Code, copy the three files of tools/vscode/ into .vscode/ once: then Terminal → Run Task → Dashboard (or Ctrl+Shift+B) starts the dashboard, Run → Dashboard (Streamlit) starts it under the debugger, and the test explorer picks up pytest.

Each notebook also has its own target (make data, make eda, …, make corrections; see the Makefile); python run_pipeline.py --only 04 runs one notebook. python -m src.advisor challenges your own decision at the keyboard.

Map of the repository

notebooks/   01 → 13, run in order by make all; each made of numbered sections with a Reading-this cell under every result
src/         the shared functions, grouped above; src/config.py holds every constant with a comment, src/paths.py every path
app/         app.py (make app, make dashboard, dashboard.bat) → decide.py, example.py, executive.py; reads results/ only
tools/       vscode/: task, launch and settings files to copy into .vscode/
prompts/     the advisor's system prompt and every prompt template it is given, verbatim
data/        raw/ (not committed; the manifest lists it) · interim/ (not committed, rebuilt) · processed/ (small final tables) · README.md
results/     every number the report cites, written by the notebooks; report_numbers.csv keys them; v8/, v9/, v10/, v2/ the registered records
figures/     every figure, PNG + SVG; figures/<notebook>/ holds each notebook's exploratory figures; figures/executive/ the report's
report/      the report (PDF and Word), the poster (PDF and typst source), figure_notes.md (every figure's caption and notebook), references.md
docs/        the course submissions, the mentor exchange, the record of decisions and the reviews the report cites; docs/README.md indexes them
tests/       pytest

Statement of work

All components of this project — data engineering, modeling, evaluation, visualization and writing — were carried out by Alan Mössinger. Code adapted from an external source is credited in-line where it appears. The ERCOT price client and the battery dispatch model are reused, with credit, from the author's own earlier projects (Forecast to Dispatch; MADS Milestone II).

References

Rand, J., Manderlink, N., Zhang, S., Talley, C., Gorman, W., Wiser, R., Seel, J., Mulvaney Kemp, J., Jeong, S., & Kahrl, F. (2025). Queued up: 2025 edition — Characteristics of power plants seeking transmission interconnection as of the end of 2024. Lawrence Berkeley National Laboratory. https://emp.lbl.gov/queues

U.S. Energy Information Administration. (2026). Preliminary monthly electric generator inventory (EIA-860M) [Data set, accessed through the EIA API v2]. https://www.eia.gov/electricity/data/eia860m/

The full reference list is in report/references.md and in the report.