Chancing Engines Landscape: Every Admissions Prediction Model vs. Our Simulation

chancing_engines_landscape.md

Chancing Engines Landscape: Every Admissions Prediction Model vs. Our Simulation

Comprehensive inventory of 50+ commercial tools and 15 academic models, compared side-by-side with our agent-based simulation. March 2026.

Table of Contents

1. Executive Summary
2. The Taxonomy
3. Tier 1: Major Consumer Chancing Engines
4. Tier 2: AI-Powered Newer Entrants
5. Tier 3: Institutional / Counselor Platforms
6. Tier 4: Academic Index Calculators
7. Tier 5: Institution-Side Yield Prediction
8. Tier 6: Academic Research Models
9. The Master Comparison Matrix
10. Feature Coverage Heat Map
11. Where Our Simulation Sits
12. Key Findings

1. Executive Summary

We identified 50+ admissions prediction tools across 6 tiers — from billion-user platforms (Naviance, CollegeVine) to single-developer GitHub repos. The landscape breaks into two fundamentally different categories:

Individual prediction tools (all commercial engines): "Given YOUR profile, what are your chances at School X?" These treat each student-college pair in isolation.

Market simulation models (our simulation + Reardon 2016 ABM): "Given ALL applicants, seat constraints, and round dynamics, what happens in the aggregate?" These model the system.

No commercial tool models hooks (legacy/athlete/donor), seat constraints, round sequencing, yield competition, or waitlist dynamics. Our simulation is the only system that combines a calibrated logistic admissions model with full market-clearing mechanics.

The closest analog is the Stanford CEPA agent-based model (Reardon et al. 2016) — same architecture, same multi-round matching — but that model uses stylized parameters while ours is calibrated to real CDS data for 55 named colleges.

2. The Taxonomy

INDIVIDUAL PREDICTION                           MARKET SIMULATION
(What are MY chances?)                          (How does the system work?)

┌─────────────────────────────┐                ┌──────────────────────────┐
│ TIER 1: Major Consumer      │                │ ACADEMIC ABMs            │
│   CollegeVine (75 factors)  │                │   Reardon 2016 (8K stu) │
│   Scoir AI (ML, round-aware)│                │   Assayed 2023 (Jordan) │
│   Niche (GPA+SAT scatter)   │                │                          │
│   PrepScholar (GPA+SAT)     │                │ OUR SIMULATION           │
│   Naviance (school scatter) │                │   55 colleges, 6 rounds  │
│   Parchment (766K outcomes) │                │   Logistic + market clear│
│   CollegeData (2K colleges) │                │   Hooks, DI, yield, WL   │
│   College Transitions       │                │   Monte Carlo validated  │
├─────────────────────────────┤                └──────────────────────────┘
│ TIER 2: AI Newcomers        │
│   20+ startups (2023-2026)  │                EMPIRICAL RESEARCH MODELS
│   GradGPT, Ultra, Chancify  │                (What factors matter?)
│   Admissionsly, DreamyUni   │                ┌──────────────────────────┐
├─────────────────────────────┤                │ Arcidiacono (SFFA trial) │
│ TIER 3: Institutional       │                │   150K apps, 370 vars    │
│   Scoir, Overgrad, Kollegio │                │   Pseudo R²=0.56         │
├─────────────────────────────┤                │ Chetty 2023 (Ivy-Plus)   │
│ TIER 4: Academic Index      │                │   ALDC decomposition     │
│   Top Tier, Ivy League Guru │                │ Espenshade (SAT equiv)   │
├─────────────────────────────┤                │   124K apps, 10 colleges │
│ TIER 5: Yield Prediction    │                │ Avery-Hoxby (revealed    │
│   ENROLL, Yield+, Slate     │                │   preference ranking)    │
│   (institution-facing only) │                └──────────────────────────┘
└─────────────────────────────┘

3. Tier 1: Major Consumer Chancing Engines

CollegeVine — The Market Leader

See collegevine.md and scoring_model_comparison.md for deep analysis.

Scoir Admission Intelligence — The Dark Horse

Scoir is the only commercial platform with round-level predictions, making it architecturally the most similar to our simulation among consumer-facing tools. However, it is institution-only (not public) and does not model hooks or market-clearing dynamics.

Niche.com — Scattergram Approach

PrepScholar — The Minimal Baseline

Naviance — School-Specific Historical Data

Parchment "My Chances"

CollegeData "College Chances"

College Transitions

4. Tier 2: AI-Powered Newer Entrants (2023–2026)

A wave of 20+ AI-powered chancing startups emerged in 2023–2026. Most are thin wrappers around LLMs or simple ML models with unverified accuracy claims.

None of these publish independent calibration data. The 98.2% and 90% claims are self-reported and should be treated skeptically. None model hooks, rounds, or market dynamics.

Ultra (Y Combinator-backed) is the most interesting — it "simulates the process that elite colleges use to evaluate profiles" by running simulated AO evaluations, which is conceptually closer to our approach than a statistical lookup.

5. Tier 3: Institutional / Counselor Platforms

Scoir stands alone as the only platform with round-level predictions (ED vs. EA vs. RD).

6. Tier 4: Academic Index Calculators

These compute the Ivy League Academic Index — a pure academic metric, not an admission probability.

Our simulation uses a similar Academic Index: AI = gpaToCSG(gpa) + section_weighted_SAT, range ~100–240. Our formula is inspired by the Ivy AI but adapted for a broader college universe.

7. Tier 5: Institution-Side Yield Prediction (Not Student-Facing)

These tools predict which admitted students will enroll, not who gets admitted. They serve colleges, not students.

Our simulation models yield from the student side — studentFinalDecisions() uses Chetty yield data by income bracket. These tools model yield from the institution side — predicting which students in an admitted class will deposit.

8. Tier 6: Academic Research Models

Arcidiacono et al. — The SFFA Harvard Model (The Gold Standard)

The most detailed empirical admissions model with publicly known coefficients.

Key Odds Ratios:

Critical comparison: Arcidiacono's athlete odds ratio (~5,075×) is far larger than our 4.5× because his model includes Harvard's internal athletic rating as a separate variable. Our model combines the athletic rating effect into the single recruited_athlete hook multiplier. Also, our multipliers operate in logit space on a pre-threshold score, while Arcidiacono's are direct odds ratios from a fully specified model.

Chetty, Deming, Friedman (2023) — "Diversifying Society's Leaders?"

Decomposition of top-1% family admissions advantage:

Our simulation captures all three channels: legacy via HOOK_MULT_BY_TIER.legacy, athletes via HOOK_MULT_BY_TIER.athlete, and non-academic ratings via ec_quality and essay_base (which correlate with structural advantage through archetype generation).

Key finding: Legacy applicants from top 1% families are 5× more likely to be admitted; recruited athletes are admitted "with near certainty." Both match our model's calibration.

Espenshade, Chung, Walling (2004) — SAT-Point Equivalents

SAT-point equivalents of admission advantages (1600 scale):

These Espenshade figures are pre-SFFA. Post-SFFA, the race advantages are eliminated or reversed in our model — a key update that no commercial chancing engine has fully incorporated.

Reardon et al. (2016) — The Agent-Based Model

The closest published analog to our simulation.

Parameter comparison:

Avery, Glickman, Hoxby, Metrick (2004/2013) — Revealed Preference Ranking

Our studentFinalDecisions() function is conceptually similar — students choose among acceptances based on utility (prestige, aid, fit). Avery-Hoxby formalizes this as a maximum likelihood estimation of college desirability parameters.

Avery and Levin (2010) — Early Admissions Game Theory

Our ED multipliers (1.2–8.0×) are the empirical realization of this theoretical prediction. Schools like Dartmouth (3.5×) and Columbia (3.4×) have massive ED advantages precisely because of this signaling equilibrium.

Hurwitz (2011) — Legacy at 30 Elite Institutions

Our HOOK_MULT_BY_TIER.legacy values (5.7× at T1, 4.0× at T2, 3.0× at T3) bracket this finding — they're calibrated higher at T1 (HYPSM) and lower at T3+.

Dale and Krueger (2002) — Does Selectivity Matter?

Not directly an admissions model, but validates our structural approach: modeling the application portfolio (list building) is as important as modeling admission probability.

UC Irvine Deep Learning (2024)

Cornell Learned Admission-Prediction (2023, Best Paper Award)

Lira et al. (2023) — AI for Personal Qualities (Science Advances)

This is the frontier: NLP models that can actually evaluate essay quality. Our simulation approximates this with essay_base (archetype-generated 1–10 score), but Lira's approach could eventually replace the proxy.

9. The Master Comparison Matrix

10. Feature Coverage Heat Map

What each model includes (filled) vs. excludes (empty):

                        CV  Scoir Niche Prep Nav  Parch CData CTrans Our  Rear  Arc  Esp  Chet
GPA                     ●    ●     ●     ●    ●    ●     ●     ●     ●    ●     ●    ●    ●
SAT/ACT                 ●    ●     ●     ●    ●    ●     ●     ●     ●    ●     ●    ●    ●
SAT section scores      ○    ○     ○     ○    ○    ○     ○     ○     ●    ○     ●    ○    ○
Course rigor (APs)      ●    ○     ○     ○    ○    ○     ●     ○     ○*   ○     ●    ○    ○
Extracurriculars        ●    ○     ○     ○    ○    ○     ●     ○     ●    ○     ●    ○    ○
Essay quality           ○    ○     ○     ○    ○    ○     ○     ○     ●    ○     ●    ○    ○
Recommendations         ○    ○     ○     ○    ○    ○     ○     ○     ○    ○     ●    ○    ○
Gender                  ●    ●     ○     ○    ○    ○     ○     ○     ●    ○     ●    ●    ○
Race/ethnicity          ◐    ●     ○     ○    ○    ○     ○     ○     ●    ○     ●    ●    ○
State/geography         ●    ●     ○     ○    ○    ○     ○     ●     ●    ○     ●    ○    ○
Intended major          ●    ○     ●     ○    ○    ○     ○     ○     ●    ○     ●    ○    ○
Legacy                  ○    ○     ○     ○    ○    ○     ○     ●     ●    ○     ●    ●    ●
Recruited athlete       ○    ○     ○     ○    ○    ○     ○     ●     ●    ○     ●    ●    ●
Donor/development       ○    ○     ○     ○    ○    ○     ○     ○     ●    ○     ●    ○    ●
First-gen               ○    ●     ○     ○    ○    ○     ○     ●     ●    ○     ●    ○    ○
Income effects          ○    ○     ○     ○    ○    ○     ○     ●     ●    ●     ○    ○    ●
Feeder school           ○    ○     ○     ○    ●**  ○     ○     ●     ●    ○     ○    ○    ○
ED/EA round             ◐    ●     ○     ○    ○    ○     ○     ○     ●    ○     ○    ○    ○
Demonstrated interest   ○    ○     ○     ○    ○    ○     ○     ○     ●    ○     ○    ○    ○
Yield protection        ○    ○     ○     ○    ○    ○     ○     ○     ●    ○     ○    ○    ○
Seat constraints        ○    ○     ○     ○    ○    ○     ○     ○     ●    ●     ○    ○    ○
Multi-round processing  ○    ○     ○     ○    ○    ○     ○     ○     ●    ○     ○    ○    ○
Yield/enrollment model  ○    ○     ○     ○    ○    ○     ○     ○     ●    ●     ○    ○    ●
Waitlist                ○    ○     ○     ○    ○    ○     ○     ○     ●    ○     ○    ○    ○
Summer melt             ○    ○     ○     ○    ○    ○     ○     ○     ●    ○     ○    ○    ○
Phantom applicants      ○    ○     ○     ○    ○    ○     ○     ○     ●    ○     ○    ○    ○
Holistic noise          ○    ○     ○     ○    ○    ○     ○     ○     ●    ●     ○    ○    ○

● = modeled   ◐ = partially   ○ = not modeled
*  AP count generated per archetype but not directly in admission score
** Naviance uses YOUR school's historical data (implicit feeder effect)

CV=CollegeVine, Prep=PrepScholar, Nav=Naviance, Parch=Parchment,
CData=CollegeData, CTrans=College Transitions, Rear=Reardon 2016,
Arc=Arcidiacono, Esp=Espenshade, Chet=Chetty 2023

11. Where Our Simulation Sits

What No One Else Does

1. Full market clearing: ED fills seats → fewer for RD. Student choices affect yield. Waitlist fills gaps. Summer melt creates final attrition. No other tool models these cascading effects.

2. Phantom applicant correction: At MODEL_SCALE = 0.013, we represent 1.3% of real applicants. Phantom applicants consume seats proportionally. No other model handles this.

3. Hook stacking in logit space: A legacy athlete donor doesn't get P > 1 because hooks are additive in log-odds. The Arcidiacono model does this too, but for a single institution; we do it across 55 colleges with tier-dependent multipliers.

4. Demonstrated interest with round interaction: DI matters most in RD (full boost), less in EA (0.5×), minimal in ED (0.2×). No tool models this.

5. Yield protection: Overqualified unhooked students get penalized at schools like Tufts and Northeastern. Only our simulation and College Transitions even acknowledge this exists.

6. Monte Carlo batch mode: 100-run headless simulations producing per-college mean ± SE for acceptance rate, yield, enrolled count, and avg SAT. No consumer tool offers this.

What Others Do Better

1. College coverage: CollegeVine covers 1,500+ schools; we cover 55. For a student considering mid-selectivity schools, CollegeVine is more useful.

2. Individual precision: CollegeVine takes YOUR actual GPA and activities. We generate synthetic students from archetype distributions. For personal prediction, CollegeVine wins.

3. School-specific history: Naviance shows where students FROM YOUR SCHOOL were admitted. We use feeder tier as a proxy but lack school-specific granularity.

4. Training data scale: Scoir has "tens of millions" of historical records; Parchment has 766K; we calibrate against CDS aggregates. Their n > our n.

Our Unique Position

                    INDIVIDUAL ←————————→ SYSTEMIC
                    PREDICTION              SIMULATION

MOST                CollegeVine
FACTORS             Scoir AI
                    Arcidiacono (370 vars)
                    CollegeData
                    College Transitions
                                            ╔════════════════╗
                                            ║ OUR SIMULATION ║
                                            ║ 30 factors +   ║
                                            ║ market clearing ║
                                            ╚════════════════╝
                                            Reardon 2016 ABM
FEWEST              PrepScholar
FACTORS             Niche

We sit at the intersection of rich factor modeling (comparable to CollegeVine's factor count, with hooks that CollegeVine can't model) and system simulation (comparable to Reardon's ABM, but calibrated to real data). No other tool occupies this position.

12. Key Findings

1. Most chancing tools are remarkably primitive

PrepScholar, Niche, CampusReel, U.S. News, and Parchment are essentially GPA + SAT lookup tables. They provide no more insight than checking the CDS yourself.

2. CollegeVine is the clear consumer leader but has structural blind spots

75 factors and 1,500+ schools, but it cannot distinguish a legacy athlete from a first-gen student with identical stats. At Harvard, the legacy athlete's admission probability is ~30× higher. This isn't a minor omission — it's the dominant factor at selective schools.

3. Scoir is the only commercial platform with round-level predictions

This makes it architecturally the most similar to our approach among commercial tools. But it's institutional-only, doesn't model hooks, and doesn't simulate market dynamics.

4. The AI startup wave is mostly noise

20+ new "AI chancing" tools launched in 2023–2026. Most are thin wrappers with unverified accuracy claims (98.2%, 90%, 85–90%). None publish calibration data. None model hooks or rounds.

5. Academic models provide the best coefficients

Arcidiacono's 370-variable Harvard model (Pseudo R² = 0.56), Espenshade's SAT-point equivalents, Chetty's ALDC decomposition, and Hurwitz's legacy odds ratios are the gold standard for calibrating hook multipliers — and they're what we used.

6. Reardon 2016 is our closest ancestor

Same ABM architecture, same multi-round matching, same focus on structural inequality. But Reardon uses stylized colleges with 150 generic seats; we use 55 named colleges with real CDS data. Reardon uses 2 student attributes (SES, caliber); we use ~30.

7. No tool models the full system

The admissions market is a multi-round, constrained matching market with information asymmetries, strategic behavior (ED signaling), supply constraints (class sizes), and cascading effects (one student's enrollment affects another school's yield). Our simulation is the only system that models all of these simultaneously.

Sources

Commercial Tools

• CollegeVine Calculator
• Scoir Admission Intelligence
• Niche Calculator
• PrepScholar Calculator
• Naviance Scattergrams
• Parchment My Chances
• CollegeData College Chances
• College Transitions Calculator
• CampusReel Calculator
• Appily Calculator
• U.S. News Calculator
• GradGPT
• CollegeAI
• AdmitYogi
• Chancify AI
• Admissionsly
• stdnt.space
• Ultra
• LumiSource
• KapAdvisor
• Top Tier Admissions Academic Index
• Ivy League Guru
• Academic Index AI

Academic Research

• Arcidiacono, Kinsler, Ransom — Asian American Discrimination (NBER WP 27068)
• Arcidiacono, Kinsler, Ransom — Legacy and Athlete Preferences (NBER WP 26316)
• Chetty, Deming, Friedman — Diversifying Society's Leaders? (NBER WP 31492)
• Opportunity Insights Non-Technical Summary
• Espenshade, Chung, Walling — Admission Preferences (SSQ 2004)
• Reardon et al. — Agent-Based Simulation (JASSS 2016)
• Avery, Glickman, Hoxby, Metrick — Revealed Preference Ranking (QJE 2013)
• Avery and Levin — Early Admissions (AER 2010)
• Hurwitz — Legacy at 30 Institutions (EER 2011)
• Dale and Krueger — Selectivity Payoff (QJE 2002, NBER WP 7322)
• Epple, Romano, Sieg — Market for Higher Education (Econometrica 2006)
• Mulhern — Naviance Impact (JLE 2021)
• Hoxby and Avery — Missing One-Offs (NBER WP 18586)
• Lira et al. — AI for Personal Qualities (Science Advances 2023)
• Cornell Learned Admission-Prediction (arXiv 2302.03610)
• UC Irvine Deep Learning Admissions (arXiv 2401.11698)
• ChanceyNN (GitHub)
• Chetty et al. — Mobility Report Cards (QJE 2020, NBER WP 23618)