A project of BYU Marriott School of Business

Data that teaches like the real world.

BlueForge is a data simulation platform for analytics and information systems courses. Faculty define a scenario, configure how the data should be wrong, and stream it to students through an authenticated API. The shares behave like production data, messy, continuous, evolving over time, so that the pipelines and analyses students build have to handle what real data actually does.

See how it works Browse use cases

How it works

Three steps from a syllabus to a stream of data.

The model is small on purpose. A faculty member should be able to take a learning objective from their syllabus and turn it into a live data source before the next class meeting.

01 / Define

Define a scenario

Describe the world the data is from. An e-commerce returns flow, a clinic's patient encounters, a manufacturing telemetry stream. BlueForge generates entities, relationships, and a stable identity for each row so the data joins like it should.

02 / Configure

Configure the messiness

Decide what should go wrong, and how often. Missingness by mechanism, late arrivals with a delay distribution, slowly-changing dimensions, drift events scheduled to specific course weeks. Every injection ships with a diagnostic and a remediation in the answer key.

03 / Stream

Stream to students

Students hit the share through an authenticated API. Everyone on the roster pulls the same live, reproducible stream, so the whole class builds against one shared realistic dataset and the same answer key. The stream keeps running for the length of the term.

Use cases

Built around the issues each course actually teaches.

The platform is one product but the priorities differ by course. Data engineering cares about late arrivals and schema drift. Statistics cares about heteroscedasticity and confounding. The shares are configured accordingly.

IS 566 · Week 5

A high-volume click and page-view stream with the failure modes a production pipeline has to survive. Students build watermarked ingestion, route bad records, and adapt when the schema changes mid-term.

Inject3% late arrivals, exp(λ=90s) with a heavy tail to 48 hours
Inject0.5% duplicate events from simulated client retries
DriftWeek 6: new field device_type appears in payload
DriftWeek 8: region encoding switches from codes to names
Inject1% of events carry an orphan user_id

See the data engineering preset →

web_events / live preview240 rowsupdated 12s ago

event_id	event_time	ingest_time	user_id	action	region
evt_8af2c1	14:02:18	14:02:19	u_4812	view	UT
evt_8af2c2	14:02:19	14:02:20	u_2207	click	CA
evt_8af2c3	14:02:20	14:48:51 +47m	u_4812	view	UT
evt_8af2c4	14:02:21	14:02:21	u_9913 orphan	view	NV
evt_8af2c5	14:02:22	14:02:23	u_2207	add_cart	CA
evt_8af2c6	14:02:23	14:02:24	u_4471	view	null
evt_8af2c7	14:02:24	14:02:25	u_4812	click	UT
evt_8af2c8	14:02:25	14:02:25	u_1180	view dup	TX

IS 414 · Week 3

A descriptive analytics dataset with a Simpson's paradox built in. The aggregate trend favors one region; disaggregated by salesperson tier, the conclusion reverses. Students learn to question aggregates before they report them.

SetupTier mix differs between regions, by design
InjectHeavy-tailed deal_size on the West cohort
InjectSlowly-changing segment attribute on customer
Inject4% MCAR missingness on discount
DriftSampling drift: a new sales territory comes online in week 9

See the analytics preset →

regional_sales / Q3 aggregateby region × tier

region	tier	deals	avg_size	win_rate
East	Senior	412	$18,420	0.38
East	Junior	128	$6,210	0.19
West	Senior	141	$22,180	0.44
West	Junior	389	$7,140	0.23
Aggregate by region
East	–	540	$15,535	0.33 aggregate
West	–	530	$11,135	0.29

IS 545 · Week 7

A binary classification share with deliberate class imbalance, concept drift, and a feature that quietly leaks the target. The point is to teach students to evaluate beyond accuracy and to find the leak before they ship the model.

SetupPositive class at 0.8%, drifting to 2.1% by week 10
InjectConcept drift: P(y | x) shifts mid-term
TrapDerived feature days_to_decision contains target leakage
InjectNew categorical level in channel after week 8
InjectHeavy-tailed requested_amount destabilizes tree splits

See the machine learning preset →

credit_decisions / class balance over time16-week run

week	records	positives	rate	notes
1	1,840	15	0.82%	baseline
2	1,902	14	0.74%	baseline
3	1,855	17	0.92%	baseline
4	1,876	22	1.17%	ramp
5	1,910	28	1.47%	ramp
6	1,889	31	1.64%	ramp
7	1,921	35	1.82%	concept drift
8	1,943	38	1.96%	concept drift

STAT 330 · Week 4

A regression dataset that fails its diagnostics in three different ways at once. Heteroscedasticity, non-normal residuals, and a single high-leverage point. Students learn that the model can fit and still be wrong.

SetupResidual variance scales with years_experience
SetupOne observation with Cook's distance > 1
SetupRight-skewed residuals (lognormal-flavored target)
InjectMulticollinearity between tenure and level (VIF ≈ 8)
Inject6% MAR missingness on bonus

See the statistics preset →

salary_survey / residual diagnosticsn = 1,200

id	fitted	residual	leverage	cook's d
obs_0012	82,140	+3,210	0.004	0.01
obs_0184	94,680	−2,140	0.006	0.00
obs_0247	118,400	+14,820	0.038	0.08 tail
obs_0398	76,210	−1,840	0.003	0.00
obs_0512	412,000	+218,400	0.182	1.34 influential
obs_0654	88,920	+940	0.004	0.00
obs_0820	132,180	+18,200	0.041	0.11 tail
obs_1041	94,140	−12,640	0.012	0.02

IS 312 · Week 6

A geographic share where the default chart misleads. Choropleth by total count makes one region look like an outlier; per-capita reverses the story. Students learn that the encoding choice is the analysis.

SetupPopulation varies 60× across cities
SetupBimodal distribution masquerading as outliers
InjectLikert-style responses that should not be averaged
SetupTime series with a regime change in week 11
Inject3% outliers from data entry errors mixed with legitimate tail

See the visualization preset →

city_metrics / raw vs. per-capita14 cities

city	population	incidents	per 10k	rank shift
Provo	115,400	182	15.8	↑ 4
Salt Lake City	204,500	412	20.1	↑ 2
Ogden	87,900	141	16.0	↑ 3
Logan	53,200	62	11.7	↑ 5
St. George	102,700	88	8.6	↓ 2
Las Vegas	651,300	1,820	28.0 leader	–
Phoenix	1,608,100	3,140	19.5	↓ 4
Boise	235,000	218	9.3	↓ 3

Institutional context

Built at BYU Marriott.

BlueForge is a project of the BYU Marriott School of Business, built by Dave Wilson, a professor of Information Systems. It exists to give business school students experience with the imperfect, in-motion data they will encounter in industry.

It is not a commercial product. There is no roadmap to commercialization, no enterprise plan, no customer logos. The platform is available to faculty inside BYU Marriott, and source code and documentation are open for colleagues to read and learn from.

Status: Phase 1 · live
Pilot term: Fall 2026 · IS 566
Author: Dave Wilson, BYU Marriott