An Interaction-Aware Decision Framework for FDIC-Assisted Distressed Bank Acquisitions: Empirical Evidence from the FDIC Failed-Bank Panel and the SVB-First Citizens Case
Authors: Yousef Nejatbakhsh, Malihe Aliasgari, Parsa Safaee, Sina Shokoohyar
Organizations: Department of Statistics, Rutgers University, New Brunswick, NJ 08901, USA · Department of Mathematical Sciences, Kean University, Union, NJ 07083, USA · Eindhoven Artificial Intelligence Systems Institute (EAISI), Eindhoven University of Technology, Eindhoven, The Netherlands · Department of Computing and Decision Sciences Stillman School of Business Seton Hall University Jubilee Hall 615 400 South Orange Avenue South Orange, NJ. 07079
Acquirers of failed banks weigh acquisition value, regulatory protection, risk, buyer strength, and execution difficulty jointly, because each signal's meaning depends on the others: high non-performing loans matter less when government loss sharing absorbs the losses. The proposed Spider Network Filtering Framework encodes such dependencies as four interpretable interaction rules among five dimensions, yielding graph-inspired features for a logistic model and a Pursue/Caution/Avoid decision rule. We evaluate it on 55 acquisitions assisted by the Federal Deposit Insurance Corporation in 2008-2024 by acquirers with a parent listed in the United States. Six financial inputs come from regulatory reports published before each failure; two deal terms, which the winning bidder knows when it bids but investors learn at the announcement, come from the regulator's press releases. The outcome, a positive three-day abnormal return around the announcement, records the market's initial reception, not realized success. Under grouped, nested cross-validation, the Spider-enhanced model reaches a mean area under the receiver operating characteristic curve of 0.594, against 0.615 for logistic regression on the same inputs and 0.632 for ridge regression; permutation tests that repeat the full model selection give at most weak evidence that any of them ranks better than chance. The Spider score used directly does not rank better than chance, and an apparent link between target distress and higher returns disappears against a bank-sector benchmark. Thresholds tuned for the F1 score approve most acquisitions with negative returns and realize negative utility when each costs three times what a positively received acquisition gains.
Figures & tables
Rule
Condition
Adj.
Protected Credit Risk ( Qi )
Ni≥N75 and Fi=1
+0.08
Recoverable Distress ( Ui )
Li=1 and Ti=0
+0.07
Structural-Risk Penalty ( Zi )
Ti=1 and Ci<C25
−0.10
Size-Burden Penalty ( Ki )
Mi>M75 and Ci<C25
−0.06
TABLE I: Spider Network Interaction Rules
Policy
Thresh.
Pursue
TP/FP
F1
FPR
Util.
4A: Fixed ( S~ )
0.70/0.40
3.1
2.9/0.2
0.151
0.011
0.041
4B: F1-tuned
0.55–0.80
40.2
27.6/12.6
0.734
0.629
− 0.184
4C: Cost (3:1)
0.75/0.45
11.8
7.9/3.9
0.332
0.196
− 0.071
Trivial policies
Pursue all
—
55
35/20
0.778
1.000
− 0.455
Pursue none
—
0
0/0
0.000
0.000
0.000
TABLE II: Model 4 Decision Policies ( n=55 ; Mean over 50 Repeats; Utility with B=1 , L=3 )
Own ratio
3:1
Ratio
Thr.
Pursue
TP/FP
FPR
F1
Util.
Gain
Util.
1:1
0.50
45.6
30.8/14.7
0.737
0.765
0.292
0.020
− 0.244
2:1
0.67
24.3
17.6/6.8
0.339
0.590
0.073
0.073
− 0.051
3:1
0.75
11.8
7.9/3.9
0.196
0.332
− 0.071
− 0.071
− 0.071
4:1
0.80
7.0
4.3/2.8
0.138
0.197
− 0.123
− 0.123
− 0.073
5:1
0.83
5.1
2.7/2.4
0.119
0.133
− 0.167
− 0.167
− 0.080
TABLE III: Utility Sensitivity to the Assumed Loss-to-Benefit Ratio (Model 3 Probabilities; Mean over 50 Repeats)
Model
Purpose
Output
Interpretability
Model 1
Decision baseline
Spider score
Very high
Model 2
Statistical benchmark
Probability
Moderate
Model 3
Hybrid predictive model
Probability
Moderate to high
Model 4
Executive decision rule
Category
Very high
TABLE IV: Comparative Summary of Four Models
Variable
Description (timing)
Network role
Median [IQR] or count
N
NPL ratio of the failed bank, noncurrent loans / loans (%) (last public call report)
Risk
13.96 [9.00, 17.23]
G
Deposits of the failed bank, billions of USD (same report)
Value
1.00 [0.50, 2.63]
L
Change in the failed bank’s deposits over the two quarters ending at that report (%)
Risk (via Li )
−4.54 [ −11.71 , 1.06]
T
Change in the failed bank’s NPL ratio over the same two quarters (percentage points)
Risk (via Ti )
2.80 [0.70, 6.56]
F
FDIC loss-share agreement (term of the bid; public with the FDIC’s announcement)
Protection
40 of 55
H
Bridge bank or systemic-risk exception (SVB only; public before its sale)
Protection
1 of 55
TABLE V: Variable Definitions, Network Roles, and Distribution in the Primary Panel ( n=55 )
Component
Raw value
Model value
Deposits (Dec. 31, 2022)
$175.4B
G=0.931
Value score V
—
0.931
FDIC loss share
Yes
F=1
Government support
Bridge bank
H=1
Protection score P
—
1.000
NPL ratio
0.19%
N=0.002
TABLE VI: Spider Network Inputs and Scores: SVB–First Citizens
Model
AUC [5th, 95th pct.]
Event-weighted AUC
Brier [5th, 95th pct.]
Brier skill
L2 logistic, C=1 (Model 2)
0.615 [0.535, 0.672]
0.551
0.238 [0.218, 0.266]
− 2.7%
LASSO logistic
0.518 [0.416, 0.620]
0.479
0.249 [0.226, 0.282]
− 7.6%
L2 logistic, C tuned (ridge)
0.632 [0.556, 0.684]
0.575
0.225 [0.211, 0.249]
2.9%
Decision tree
0.585 [0.448, 0.675]
0.533
0.291 [0.240, 0.362]
− 25.9%
Random forest
0.605 [0.545, 0.666]
0.569
0.231 [0.216, 0.244]
0.1%
XGBoost
0.604 [0.534, 0.656]
0.573
0.243 [0.223, 0.267]
− 5.0%
TABLE VII: Model Comparison: 50 Repeats of Grouped 5-Fold Cross-Validation ( n=55 , 51 Event Clusters)
50 partitions
Cluster bootstrap, AUC diff. [95% interval]
Comparison
AUC diff. [5th, 95th pct.]
M3 higher
Brier diff. [5th, 95th pct.]
Conditional
Refitting
M3 − Model 2 ( C=1 )
− 0.020 [ − 0.085, + 0.016]
28%
− 0.006 [ − 0.024, + 0.017]
− 0.044 [ − 0.104, + 0.001]
[ − 0.155, + 0.131]
M3 − Ridge ( C tuned)
− 0.038 [ − 0.092, + 0.010]
10%
+ 0.008 [ − 0.009, + 0.028]
− 0.056 [ − 0.115, − 0.010]
[ − 0.166, + 0.117]
M3 − LR + pairwise
− 0.038 [ − 0.085, + 0.032]
22%
+ 0.017 [ − 0.004, + 0.037]
− 0.070 [ − 0.214, + 0.049]
—
M3 − Random forest
− 0.011 [ − 0.079, + 0.076]
44%
+ 0.001 [ − 0.019, + 0.027]
− 0.050 [ − 0.188, + 0.066]
—
M3 − XGBoost
− 0.010 [ − 0.083, + 0.055]
42%
− 0.011 [ − 0.031, + 0.010]
—
—
TABLE VIII: Paired Comparison of Model 3 with Selected Baselines (Same Partitions, 50 Repeats)
Null AUC
p -value
Model
AUC
Mean
95th
1-sided
Holm
2-sided
Model 2 ( C=1 )
0.635
0.479
0.659
0.08
0.16
0.24
Ridge ( C tuned)
0.668
0.478
0.653
0.04
0.13
0.14
Model 3 ( C tuned)
0.600
0.477
0.647
0.12
0.16
0.36
Spider score (Model 1)
0.421
0.499
0.630
0.82
—
0.36
TABLE IX: Permutation Tests: Is Discrimination Above Chance?
Stage
AUC
Event AUC
Brier
A: measured inputs
0.632
0.575
0.225
B: A + Spider score
0.630
0.573
0.225
C: A + interaction rules
0.598
0.534
0.231
D: full SNFF (Model 3)
0.594
0.529
0.232
E: D without protection inputs
0.615
0.556
0.228
F: D without buyer capital
0.632
0.573
0.225
TABLE X: Ablation Study: 50 Repeats of Grouped 5-Fold Cross-Validation
Variant
Positive CAR
Model 2
Ridge
Model 3
Spider score
Primary specification
35/55
0.615
0.632
0.594
0.423
Outcome: CAR[0,+1] around announcement date
39/55
0.441
0.508
0.507
0.473
Outcome: CAR[−1,+1] around announcement date
41/55
0.572
0.601
0.591
0.425
Outcome: market-model CAR
31/55
0.642
0.648
0.647
0.409
Outcome: CAR against bank-sector index
39/55
0.277
0.280
0.413
0.505
Published inputs only (no F , H )
35/55
0.635
0.646
0.615
0.466
TABLE XI: Robustness: Mean AUC over 50 Repeats under Alternative Outcomes, Specifications, and Sample
Primary panel (55)
Other 35 deals
Component
Sign in S
AUC
p
AUC
p
Spider score S~
+
0.414
0.35
0.507
0.96
Value V
+
0.533
0.72
0.464
0.31
Protection P
+
0.454
0.54
0.464
0.71
Buyer strength B
+
0.537
0.66
0.714
0.06
Risk R
−
0.699
0.04
0.724
0.04
TABLE XII: Direction of the Spider Score and Its Components (AUC against a Positive CAR)
A. Benchmark and window: AUC (two-sided p )
Outcome
Primary 55
Other 35
SPY, reaction day (primary)
0.699 (0.04)
0.724 (0.04)
Bank-sector index, reaction day
0.514 (0.90)
0.561 (0.66)
Market model, reaction day
0.644 (0.14)
0.637 (0.22)
SPY, [0,+1] around release date
0.423 (0.47)
0.572 (0.57)
SPY, [−1,+1] around release date
0.343 (0.13)
0.562 (0.62)
TABLE XIII: Sensitivity of the Risk Component’s Association with Announcement Returns
User
Use of the framework
Acquiring bank
Structuring the review of a distressed target
FDIC / regulators
Comparing bidders and transaction structures
Risk committee
Auditing the drivers of a recommendation
Board / executives
Pursue, Caution, or Avoid deliberations
Investors
Assessing the sources of acquisition gains
TABLE XIV: Potential Users of an Interaction-Aware Decision Framework
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Failed bank
Parent
Announced
FDIC release
Timing
LS
H
CAR (%)
Y
Washington Mutual Bank
JPM
2008-09-25
PR-85-2008
A
0
0
9.52
1
First Republic Bank
JPM
2023-05-01
PR-34-2023
P
1
0
1.77
1
Silicon Valley Bank
FCNCA
2023-03-26
PR-23-2023
N
1
1
54.31
1
Colonial Bank
BBT
2009-08-14
PR-143-2009
A
1
0
− 3.85
0
Downey Savings and Loan Assoc., F.A.
USB
2008-11-21
PR-124-2008
A
1
0
0.71
1
Westernbank Puerto Rico
BPOP
2010-04-30
PR-97-2010
A
1
0
3.33
1
Appendix
TABLE XV: Primary Panel: Source-Verified Terms, Timing, and Outcome of the 55 Transactions
The Spatial-Temporal Graph Attention Network (ST-GAT) framework was created to serve as an explainable GNN-based solution for detecting bank distress early warning signs and for conducting macro-prudential surveillance of the interbank system in the United States. The ST-GAT framework models 8,103 FDIC insured institutions across 58 quarterly snapshots (2010Q1-2024Q2). Bilateral exposures were reconstructed from publicly available FDIC Call Reports using maximum entropy estimation to produce a dynamic directed weighted graph. The framework achieves the highest AUPRC among all GNN architectures (0.939 +/- 0.010), trailing only XGBoost (0.944). Ablation analysis confirms the BiLSTM temporal component contributes +0.020 AUPRC; temporal attention weights exhibit a monotonically decreasing pattern consistent with long-run structural vulnerability weighting. Permutation importance identifies ROA (0.309) and NPL Ratio (0.252) as dominant predictors, consistent with post-mortem analyses of the 2023 regional banking crisis. All data are publicly available FDIC Call Reports and FRED series; all code and results are released.
Mohammad Nasir Uddin
Data Analytics and Applied AI Researcher, Westcliff University, Irvine, CA, USA
Algorithmic recourse promises to help applicants rejected by automated systems by explaining the changes needed to secure acceptance. What incentive do decision-makers, such as banks and employers, have to offer recourse? We study this question in a screening model in which recourse is both productive and selective: completing recourse improves an applicant's value to the decision-maker, but applicants differ in their cost of completion. The optimal policy is a threshold rule: reject applicants with low scores, offer recourse to an intermediate range of scores, and accept applicants with high scores outright. Because the intermediate range spans the cutoff that would separate acceptance from rejection when recourse is not available, some marginal applicants gain a new path to acceptance, while others---who would have been accepted outright---must now clear a costly hurdle.
Matthew Olckers, Toby Walsh
Centre for Household Finance, Stellenbosch University · UNSW Sydney
The U.S. financial system processes approximately 1.3 million interbank transactions daily, yet no system in the reviewed literature models fraud propagation across the interbank network using fraud co-occurrence edge features. Prior interbank GNN architectures model credit contagion using credit distress supervision signals, producing systems misaligned for fraud forensics. No existing system generates SAR narratives with per-assertion forensic traceability to specific numerical detection outputs, creating regulatory auditability gaps in FinCEN-submitted reports. This paper introduces SCAFDS (Systemic Contagion-Aware Fraud Detection System), a seven-stage integrated surveillance pipeline addressing five structural limitations of prior art: (1) fraud-specific interbank topology encoding using fraud co-occurrence frequency metrics f(u,v,t) derived from FinCEN SAR registry records; (2) edge-feature-informed graph attention where coefficients are computed from both node representations and fraud co-occurrence edge features; (3) bilinear fraud co-occurrence risk fusion producing institution-level systemic fraud risk scores; (4) attribution-conditioned SAR narrative generation with per-assertion significance thresholds ensuring each FinCEN SAR assertion is traceable to a specific numerical pipeline output; and (5) topology-aware adaptive forensic feedback updating graph attention weights from regulatory dispositions. Experiments on the IEEE-CIS Fraud Detection Dataset (590,540 transactions) and a synthetic FDIC-aligned interbank network (8,103 institutions, 169,800 edges) show SCAFDS achieves AUPRC=0.515+/-0.032 and AUROC=0.802+/-0.018, representing +15.9pp and +13.7pp improvements over GraphSAGE-AML. Partial validation on FDIC enforcement action records (n=4,279) confirms consistent model ranking. USPTO Provisional Patent Application No. 64/061,083, filed May 8, 2026.