MaBoost: Hybrid Compliance Detection

• The paper introduces MaBoost as a hybrid framework combining a Mamba-based selective state space encoder with an XGBoost classifier for detecting insider filing delays.
• It leverages structured trade attributes and sequential event encoding to achieve high precision (up to 99.47% F1) under various regulatory settings.
• MaBoost embodies a dual concept, serving as a domain-specific detection tool and reflecting broader margin-aware boosting ideas in ensemble methods.

Searching arXiv for recent and directly relevant papers on “MaBoost” and related usage. MaBoost most commonly denotes the hybrid framework introduced for insider filing violation detection in the Insider Filing Delay (IFD) benchmark, where a Mamba-based selective state space encoder produces compact behavioral representations from sequential Form 4 events and an XGBoost classifier performs binary detection of disclosure delay risk (Huang et al., 27 Jul 2025). In a separate and older line of literature, “MaBoost” can also function as a generic shorthand for margin-aware boosting, especially in discussions surrounding margin-distribution methods such as MDBoost; however, that usage is terminological rather than canonical, because the MDBoost paper itself introduces “margin-distribution boosting” rather than an algorithm formally named MaBoost (0904.2037). The term therefore has a dual meaning: a specific financial-compliance architecture in recent arXiv work, and a broader conceptual label in the boosting literature.

1. Regulatory task and problem formulation

In the IFD framework, MaBoost is defined against a specific compliance rule in U.S. public markets: corporate insiders must report trades through SEC Form 4 within two business days after trade execution, a requirement strengthened by the Sarbanes–Oxley Act in 2002 (Huang et al., 27 Jul 2025). A filing violation is operationalized when the SEC receipt date, SECDATE, is beyond the two-business-day deadline computed from the transaction date, TRANDATE, using the SEC business calendar. The resulting machine-learning task is supervised binary classification: given a transaction with structured attributes, predict delay = 1 for a violation and delay = 0 for a compliant filing.

The paper evaluates MaBoost under three constrained regulatory settings intended to emphasize different signal regimes. In Equal Weight, all 52 structured attributes are treated uniformly. In Constraint Condition, empirically low-impact groups such as spatiotemporal features receive reduced weights, and the model must recover interactions among stronger feature groups such as governance and trade characteristics. In Suspected Violation, the emphasis shifts toward weak but correlated cues, including trading history and subtle delay signals under unusual firm or market conditions. This design places the model in a regime closer to regulatory triage than to unconstrained tabular classification.

A further distinction in the benchmark is the violation taxonomy. The binary label is delay ∈ {0,1}, with 1 assigned when SECDATE > SEC_Business_Day_Lag2. Beyond this primary label, the paper distinguishes oversight violations—approximately 77%, characterized by delay of at most three business days and infrequent violators—from intentional violations—approximately 23%, characterized by delay of at least four business days and insiders who violate at least 95% of the time. This suggests that MaBoost is intended not merely to detect lateness, but to operate in a behavioral regime where repeated or strategic noncompliance is relevant.

2. Dataset foundation and feature space

MaBoost is inseparable from IFD, the dataset on which it is defined and evaluated. IFD comprises 4,051,143 Form 4 transactions from 2002–2025, covering 7,633 firms and 15,573 insiders, with trade-level granularity retained (Huang et al., 27 Jul 2025). The dataset includes only open market transactions with TRANCODE ∈ {P, S}, excludes amended filings and option-related transactions, retains valid WRDS cleanse codes (R, H, C, L, I), and removes filings whose transaction dates occur after the report date. Two-business-day compliance labels are produced through a Python SEC business calendar covering 2002–2025.

The feature space is explicitly structured around insider behavior, governance, firm characteristics, transaction metadata, and timing fields. Role variables ROLECODE1–4 distinguish CEOs, corporate suite, beneficial owners, and other insiders. Governance variables include Blockholder Ratio and ownership concentration measured by HHI. Firm-level indicators include [ROA](https://www.emergentmind.com/topics/rectangular-occlusion-attacks-roa), Leverage, B/M, R&D/Assets, and Tobin’s Q. Transaction-level metadata include TRANCODE, adjusted shares and prices, ownership type, CLEANSE, and amendment-related fields. Timing variables include FDATE, SECDATE, SIGDATE, SEC_Business_Day, and SEC_Business_Day_Lag2, from which gap_days and delay are computed.

Several derived variables formalize historical and structural behavior:

$$InsiderRatio = \frac{Violations}{Trades}, \qquad FirmRatio = \frac{Violations\_Firm}{Trades\_Firm}$$

$$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$

$$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$

$$ROA = \frac{Net\_Income}{Total\_Assets}, \qquad Leverage = \frac{Total\_Liabilities}{Total\_Assets}$$

$$B/M = \frac{Book\_Value\_of\_Equity}{Market\_Value\_of\_Equity}, \qquad R$$

$$Tobin's\ Q = \frac{Market\_Value\_of\_Firm}{Replacement\_Value\_of\_Assets}$$

The descriptive statistics reported for IFD establish the behavioral heterogeneity that MaBoost is designed to exploit. Beneficial owners have the highest violation rate at 22.24%, followed by other insiders at 16.59%, corporate suite at 10.63%, and CEOs at 10.88%. The dataset also contains more than 56,000 stealth sequences, 697 round trips, about 43,000 trades executed within 60 days before earnings but disclosed after, and 21 trades during restatement periods. Average delays are 37 business days; intentional purchases delay about 116 days, and intentional sales about 77 days. Late-filed purchases earn about 0.03% abnormal return per day, or 4.64% total during the delay window. These observations are descriptive rather than causal, but they motivate a model family that can integrate longitudinal insider behavior with structured firm context.

3. Architecture and learning objective

MaBoost is a hybrid selective state space encoder plus gradient-boosted trees. Its first stage is a Mamba-based encoder that processes per-insider transaction timelines up to seq_len = 100; its second stage is an XGBoost classifier that uses the learned behavioral representation for interpretable binary detection (Huang et al., 27 Jul 2025). The sequential modality consists of recent Form 4 trades in temporal order, with each time step carrying transaction metadata, timing information, compliance targets, and governance and firm indicators aligned to transaction dates. Categorical variables are encoded as short one-hots or learned embeddings; continuous variables are winsorized and standardized; missing values are handled with sentinel codes and masks.

The state-space backbone uses the discrete-time selective SSM

$h_t = A_t h_{t-1} + B_t x_t, \qquad z_t = C_t h_t$

where $x_t$ is the feature vector for the $t$-th event and $A_t, B_t, C_t$ are input-dependent parameters realized through learned gating and convolutional projections. Mamba implements efficient selective scanning with linear-time complexity in sequence length. The encoder output is then aggregated into a fixed-size representation,

$$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$0

using mean pooling or the final state. In the reported MaBoost configuration, bidirectional scanning is enabled and GELU activation is applied.

The XGBoost stage models detection as an ensemble of $$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$1 trees with binary logistic objective:

$$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$2

with regularized loss

$$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$3

$$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$4

Tree boosting uses second-order Taylor approximations and a shrinkage parameter eta to control update size. The reported hyperparameters are highly specific. For the Mamba encoder, Optuna tuning yields d_model = 256, n_layers = 4, ssm_rank = 4, dropout = 0.1, activation = GELU, prenorm = True, bidirectional = True, seq_len = 100, and norm_epsilon = 5e−5; convolutional input projection and residual connections are retained. For XGBoost, the paper reports objective = binary_logistic, tree_method = hist, max_depth = 10, min_child_weight = 10, gamma = 0.8, subsample = 0.8, colsample_bytree = 0.8, learning_rate = 0.1, n_estimators = 3000, reg_alpha = 0.1, and reg_lambda = 1.0.

The end-to-end pipeline begins with WRDS Form 4 extraction and filtering, joins the SEC business calendar to generate the two-business-day deadline, constructs chronological per-insider sequences, normalizes continuous fields and encodes categorical variables, runs the Mamba encoder to obtain $$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$5, and then classifies with XGBoost to produce $$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$6 and $$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$7. Threshold $$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$8 may be calibrated to emphasize either precision or recall, depending on regulatory priorities.

4. Evaluation, baselines, and ablations

The evaluation protocol uses stratified ten-fold cross validation, reflecting both class imbalance and the need for robust comparison across constrained settings (Huang et al., 27 Jul 2025). The reported metrics are

$$TradeValue = \frac{Trade\_Amount}{Market\_Cap}, \qquad Ln(Distance) = \ln(Distance\_to\_nearest\_top10\_metro)$$9

$$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$0

Accuracy, ROC-AUC, and PR-AUC can be computed, but the principal reported results focus on Precision, Recall, and F1.

The headline results show that MaBoost attains its strongest performance under the Constraint Condition setting, where it achieves F1 up to 99.47%, with Precision 99.09% and Recall 99.85%. Under Equal Weight, it reaches F1 94.79% with Precision 94.93% and Recall 95.48%. Under Suspected Violation, it reaches F1 96.76% with Precision 96.24% and Recall 97.08%. The paper characterizes these results as showing tightly coupled precision and recall across settings, which it interprets as evidence of robust decision boundaries.

The comparative results situate MaBoost against statistical, deep learning, hybrid, and LLM baselines.

Model or paradigm	Setting	Reported result
Logistic Regression	Constraint Condition	F1 60.79%
XGBoost alone	Constraint Condition	F1 97.65%
Vanilla Mamba encoder	Constraint Condition	F1 98.69%
Bi-LSTM	Constraint Condition	F1 97.33%
Transformer	Constraint Condition	Recall 99.81%
ConViT-Transformer	Suspected Violation	F1 94.87%
GPT-4O fine-tuned	Equal Weight	F1 87.44
LLaMA-3.1-405B fine-tuned	Constraint Condition	F1 90.67
DeepSeek-V3 fine-tuned	Suspected Violation	F1 89.13

These comparisons indicate that MaBoost improves over XGBoost alone and over a standalone Mamba encoder in the reported regulatory setting. A plausible implication is that the hybridization is not merely a convenience architecture: it appears to combine long-range sequential summarization with a tree-based decision surface that remains strong on structured financial data.

The ablation studies reinforce that interpretation. Feature group removals uniformly reduce performance; under Suspected Violation, removing Insider History or Spatiotemporal features causes large F1 reductions, reported as drops to approximately 64–92, while Financial Health and Governance are especially influential in Equal Weight and Constraint Condition. Component ablations further show that removing the convolutional input projection or residual connections in Mamba degrades F1 noticeably. Variants with reduced state dimension (d8), modified time step resolution (dt1), or alternate activation (GELU) produce smaller changes, which the paper takes as confirmation of the utility of the default Mamba configuration.

5. Interpretability, surveillance use, and engineering considerations

Although MaBoost contains a sequence encoder, its interpretability is delivered primarily through the XGBoost head. The paper states that feature importance by gain or cover, together with partial dependence analysis, identifies high-risk patterns consistent with regulatory reasoning (Huang et al., 27 Jul 2025). Salient variables include insider roles, especially the elevated violation propensity of beneficial owners; timing variables such as gap_days, SECDATE beyond SEC_Business_Day_Lag2, and signature-to-filing discrepancies; governance variables such as BlockholderRatio and HHI; insider history variables InsiderRatio and FirmRatio; and trade characteristics such as large TradeValue, absence of plan-based Rule 10b5-1 trading, and purchases or sales clustered around earnings dates.

The practical interpretation of these signals is cautious. High BlockholderRatio or high HHI can correlate with reduced violation likelihood, whereas weak governance correlates with elevated risk. Stealth sequences and round-trip trades often coincide with delays, and pre-earnings trades disclosed after announcement align with potential misuse of private information. These are correlational patterns extracted from the benchmark rather than formal legal findings, but they fit the paper’s stated objective of regulatory forensics.

From a deployment perspective, the framework is designed for surveillance pipelines. Mamba’s linear-time encoder supports efficient processing for sequences of up to 100 events, while XGBoost inference remains low latency at scale. The paper describes a deployable pipeline of WRDS ETL → filters and calendar join → feature normalization/embedding → encoder inference → tree-based classification, with all code and data released publicly. Threshold calibration is explicitly part of the design: $$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$1 can be tuned for higher recall in enforcement triage or higher precision to reduce false alarms, and Platt scaling or isotonic regression can be used to calibrate $$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$2.

The reported engineering stack consists of 4×4090-Ti GPUs (24GB each), 64GB RAM, and 8 CPU cores per node, across 6 nodes. The training regime uses 200–500 epochs for the encoder with ten-fold cross validation, and hyperparameters are selected by Optuna. The paper also recommends drift monitoring over time, particularly as governance norms change, and notes that logging model decisions and feature attributions supports audit trails.

6. Terminological ambiguity and relation to margin-distribution boosting

A recurring source of confusion is that “MaBoost” can also be used informally to denote margin-aware boosting, a family of methods that explicitly optimize properties of margin distributions rather than only a convex surrogate loss. In that sense, MDBoost—introduced by Shen and Li—can be described as a specific instance of margin-aware boosting, but the paper itself does not define or brand the algorithm as “MaBoost” (0904.2037). This distinction matters because the 2025 MaBoost for insider filing detection is not a classical boosting algorithm in the AdaBoost or LPBoost sense; it is a hybrid Mamba-plus-XGBoost architecture.

MDBoost belongs to the ensemble-learning tradition built around the margin of a training example. For an ensemble decision function

$$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$3

with weak learners $$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$4 and weights $$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$5, the margin of example $$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$6 is

$$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$7

The algorithm optimizes the empirical margin distribution by maximizing average margin and minimizing margin variance simultaneously. The corresponding quadratic program is

$$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$8

or, in normalized form,

$$BlockholderRatio = \frac{Shares\_Blockholders}{Shares}, \qquad HHI = \sum_{k=1}^{n} Ownership\_Share_k^2$$9

Its dual has the LPBoost-style constraint family

$$ROA = \frac{Net\_Income}{Total\_Assets}, \qquad Leverage = \frac{Total\_Liabilities}{Total\_Assets}$$0

together with a quadratic regularizer,

$$ROA = \frac{Net\_Income}{Total\_Assets}, \qquad Leverage = \frac{Total\_Liabilities}{Total\_Assets}$$1

and is solved by totally-corrective column generation. The weak-learner oracle is

$$ROA = \frac{Net\_Income}{Total\_Assets}, \qquad Leverage = \frac{Total\_Liabilities}{Total\_Assets}$$2

Conceptually, the relationship between the two meanings of MaBoost is limited to the word “boost.” The financial-compliance MaBoost uses XGBoost as a classifier head but does not inherit MDBoost’s margin-distribution objective, column-generation procedure, or totally-corrective update rule. Conversely, MDBoost has no connection to insider filing surveillance, Mamba state-space models, or sequential event encoding. The ambiguity is therefore terminological rather than methodological.

MDBoost’s empirical evidence, reported on 13 UCI benchmarks with decision stumps, shows best test error on 9 out of 13 datasets, including twonorm 3.5 ± 0.2%, ringnorm 5.1 ± 0.4%, splice 8.2 ± 1.0%, and heart 16.1 ± 4.2%, with AdaBoost slightly better on thyroid at about 7.2% versus MDBoost at about 7.6%. These results support the margin-distribution view of boosting, but they should not be conflated with the IFD paper’s MaBoost, which addresses a different domain, architecture, and optimization target.

7. Limitations and prospective directions

The IFD paper identifies several limitations of MaBoost as a regulatory detection system (Huang et al., 27 Jul 2025). The labels depend on public filings and business-calendar computations, which may introduce edge-case errors. Amended records are excluded, but residual inaccuracies may persist. The benchmark is U.S.-specific, centered on SEC Form 4 and the SEC business calendar, so extending the approach to other jurisdictions would require domain adaptation. Potential confounders include firm size, sector, and geography as proxied by Ln(Distance), and the paper recommends monitoring possible bias across roles or firms.

The stated future directions are threefold: expansion of IFD internationally and to multilingual filings; integration of unstructured disclosures such as 10-K narratives through document embeddings; and causal learning frameworks intended to move beyond correlation toward compliance logic and proactive intervention design. These directions suggest a broader research program in which MaBoost functions as a starting point for behavior-rich, reproducible financial-compliance modeling rather than a closed solution.

In the older margin-aware boosting lineage, the limitations are different. MDBoost relies on a Gaussian-margin motivation for its interpretation of AdaBoost, requires numerical care because the matrix $$ROA = \frac{Net\_Income}{Total\_Assets}, \qquad Leverage = \frac{Total\_Liabilities}{Total\_Assets}$$3 is semidefinite and singular, and incurs higher computational cost than stage-wise AdaBoost because each iteration solves a convex quadratic program (0904.2037). Those constraints belong to the theory and optimization of margin-distribution boosting rather than to the financial-compliance system.

Taken together, the term “MaBoost” now spans two distinct research contexts. In contemporary financial surveillance, it denotes a Mamba–XGBoost framework for insider filing delay detection on the IFD benchmark. In the boosting literature, it may serve informally as shorthand for margin-aware boosting, with MDBoost as a representative example. Precision in usage is therefore essential: the former is a domain-specific hybrid classifier for SEC compliance analytics, whereas the latter is a margin-distribution optimization paradigm in ensemble learning.