Insider Filing Delay (IFD)

• Insider Filing Delay (IFD) is a regulatory measure defining delayed filing of insider trades beyond the SEC’s 2-business-day deadline.
• It operationalizes compliance violations using rule-based data processing and binary labeling based on transaction-to-filing delays.
• Hybrid models like MaBoost, combining time-series encoding and XGBoost, effectively detect IFD, influencing market transparency and regulatory forensics.

Insider Filing Delay (IFD) denotes a class of disclosure-timing failures in insider-trading regulation. In the narrower formulation introduced by the benchmark paper on SEC Form 4 compliance, IFD is the delayed disclosure of insider trades in violation of the SEC’s timing rules, chiefly the two-business-day rule introduced by the Sarbanes–Oxley Act; in that formulation, an IFD violation is any Form 4 transaction whose reporting date is later than the legally mandated deadline, operationalized as a binary label for delayed versus timely filing. In a broader formulation developed for the Form 144/Form 4 regime, IFD also includes cases in which economically relevant information reaches the market too late, in the wrong order, or without verifiable resolution, as in reporting inversion and predictive decoupling between notice-of-intent and execution disclosures (Huang et al., 27 Jul 2025, Neupane, 19 Feb 2026).

1. Regulatory definition and conceptual scope

In the Form 4 setting, IFD is tied to the SEC’s reporting framework for changes in beneficial ownership. Form 4 is the filing through which corporate insiders report changes in beneficial ownership of company securities. The relevant insider set includes officers, directors, and beneficial owners of more than 10% of a registered equity class. Under SOX (2002), Rule 16a-3 and related regulations require insider trades disclosed on Form 4 to be filed within two business days of the transaction date. The benchmark paper operationalizes this legal rule with an SEC business-day calendar for 2002–2025 and defines a violation when the actual filing or receipt date exceeds SEC_Business_Day_Lag2, the transaction date plus two SEC business days. The binary label is Delay or Id, with 1 denoting a delayed filing and 0 a timely filing (Huang et al., 27 Jul 2025).

The same paper treats IFD as both a behavioral phenomenon and a benchmark object. As a behavioral phenomenon, it encompasses strategic or negligent late reporting. As a benchmark, it refers to a labeled dataset and evaluation framework for detecting such violations. Delay is measured through timing fields including TRANDATE, FDATE, SECDATE, SEC_Business_Day, and SEC_Business_Day_Lag2, with Gap_Days measuring transaction-to-filing delay. For history-based variables, the paper defines

$Insider\ Ratio = \frac{Violations}{Trades}$

and

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

where the numerators count prior filing delays at the insider and firm levels, respectively (Huang et al., 27 Jul 2025).

The broader Form 144/Form 4 analysis extends the concept beyond late timestamps. Form 144 is a notice of proposed sale that opens a statutory 90-day window within which the proposed shares may be sold; if there is no execution within that window, the filing lapses and there is no mandatory follow-up stating that the sale did not occur. The paper defines reporting inversion as the systematic reversal in which the economically decisive event, execution reported on Form 4, is disclosed before, or without, timely disclosure of prior intent on Form 144. Citing Franzen, Li, and Vargus (2013), it reports that pre-SOX Form 144 preceded Form 4 in 93.7% of cases, whereas post-SOX that falls to 17.5%. This leads to predictive decoupling: Form 144 is supposed to signal what may happen, but it can be filed late or ex post, may never produce a trade within the 90-day window, and is never formally confirmed or negated. In that sense, IFD becomes a two-dimensional phenomenon involving delayed revelation and structural non-verifiability (Neupane, 19 Feb 2026).

2. Dataset construction and operationalization in the Form 4 benchmark

The IFD benchmark is constructed from WRDS insider trading / Form 4 data processed in SAS and covers SEC Form 4 filings for open-market purchases and sales from 2002 to 2025. The abstract describes “over one million Form 4 transactions,” while the paper also states explicitly that IFD contains 4,051,143 transactions involving over 7,633 firms and 15,573 insiders. The retained transaction types are TRANCODE = P or S, so the benchmark focuses on open-market purchases and sales; option-related trades and equity awards are excluded (Huang et al., 27 Jul 2025).

Preprocessing is rule-based and legally grounded. The paper retains only trades with valid cleanse codes R, H, C, L, I, removes amended filings via the AMEND flag, excludes option-related transactions and equity awards, excludes filings where transaction dates are after the report date, and filters out trades with abnormal prices more than ±20% from CRSP closing price. Financial variables are lagged and winsorized at the 1st and 99th percentiles. For deadline construction, each trade is merged with an LLM-generated SEC business calendar for 2002–2025, and the two-business-day deadline is stored as SEC_Business_Day_Lag2. Unlike prior work that aggregates to insider-day, the dataset preserves trade-level granularity without insider-day aggregation to support machine learning tasks (Huang et al., 27 Jul 2025).

Each sample has 52 attributes grouped into six categories: basic identification and date information; company and security information; form and transaction details; transaction codes and data cleansing; timing and delay information; and additional fields. The timing block includes FDATE, CDATE, MAINTDATE, SECDATE, SIGDATE, Gap_Days, SEC_Business_Day, SEC_Business_Day_Lag2, and Delay. The paper notes that Delay is used context-dependently as reporting lag in days or as a binary flag, while Id is described as an internal unique identifier and also used as the label, with 1 = illegal operation. Beyond raw fields, the benchmark constructs derived variables for trade severity, governance, and firm condition, including

$$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$

$$Blockholder\ Ratio = \frac{Shares\_Blockholders}{Shares}, \quad HHI = \sum_{k=1}^{n} (Ownership\_Share_k)^2,$$

and firm-level indicators such as $ROA$, Leverage, $B/M$, $R$, and Tobin’s $Q$ (Huang et al., 27 Jul 2025).

The main labeling rule is strictly legal rather than anomaly-based. For each transaction, timeliness is assessed relative to SEC_Business_Day_Lag2; if the relevant filing or SEC date passes this deadline, the sample is labeled as a violation. The paper also defines descriptive subtypes. “Oversight violations” are delayed by ≤ 3 business days and committed by insiders who do not frequently violate; “intentional violations” are delays ≥ 4 business days by insiders who violate at least 95% of the time. These are used descriptively and in feature engineering rather than as separate benchmark labels (Huang et al., 27 Jul 2025).

3. Benchmark task, model architecture, and evaluation protocol

The central task is insider filing violation detection as a binary classification problem grounded in regulatory compliance. For each transaction $i$, the label $y_i \in \{0,1\}$ indicates whether the filing is delayed. The paper evaluates three feature-configuration modes. In Equal Weight, all 52 attributes are treated uniformly. In Constraint Condition, empirical domain knowledge is introduced and some feature groups, such as spatiotemporal variables, are downweighted when they have limited standalone predictive power. In Suspected Violation, the focus shifts to subtle behavioral cues, especially trading history, delay patterns, and their alignment with abnormal financial patterns (Huang et al., 27 Jul 2025).

The proposed architecture, MaBoost, is a hybrid framework that combines a Mamba-based state space encoder with XGBoost. For insider $FirmRatio = \frac{Violations\_Firm}{Trades\_Firm},$0, a historical transaction sequence $FirmRatio = \frac{Violations\_Firm}{Trades\_Firm},$1 is encoded through the linear state-space update

$FirmRatio = \frac{Violations\_Firm}{Trades\_Firm},$2

and the outputs are aggregated into a fixed-length embedding

$FirmRatio = \frac{Violations\_Firm}{Trades\_Firm},$3

The paper states that aggregation can be mean pooling or taking the final state and that the implementation adopts a lightweight bidirectional encoder with GELU activation. Reported Mamba hyperparameters are d_model = 256, n_layers = 4, ssm_rank = 4, dropout = 0.1, use_bidirectional = True, and sequence length 100 (Huang et al., 27 Jul 2025).

The XGBoost stage consumes the learned embedding and outputs the violation probability through an additive boosted-tree model,

$FirmRatio = \frac{Violations\_Firm}{Trades\_Firm},$4

with

$FirmRatio = \frac{Violations\_Firm}{Trades\_Firm},$5

trained using regularized logistic loss

$FirmRatio = \frac{Violations\_Firm}{Trades\_Firm},$6

Reported XGBoost hyperparameters are objective = binary_logistic, max_depth = 10, n_estimators = 3000, reg_alpha = 0.1, reg_lambda = 1.0, gamma = 0.8, subsample = 0.8, colsample_bytree = 0.8, and learning_rate = 0.1. The paper characterizes this composition as combining Mamba’s strength in long-sequence time-series modeling with XGBoost’s interpretability and robustness for tabular classification (Huang et al., 27 Jul 2025).

Evaluation uses Precision, Recall, F1-Score (reported ×100%), with 10-fold cross-validation for training and hyperparameter tuning, and 200 or 500 epochs depending on the experiment. Baselines span classical models, deep sequence models, hybrid CV/sequence architectures, BERT-family models, and LLMs. The benchmark suite includes Linear Regression, Logistic Regression, Decision Tree, Random Forest, XGBoost; Seq2Seq, RNN, Bi-RNN, LSTM, Bi-LSTM, GRU, Transformer, Mamba; CNN- and ViT-based hybrids; BERT, CINO, RoBERTa; and LLM families including GPT, LLaMA 3.1, Qwen 2.5, DeepSeek, Claude 3.5, and Gemini 1.5 under zero-/few-shot prompting, embedding-based classification, and fine-tuning (Huang et al., 27 Jul 2025).

The reported MaBoost results are as follows:

Setting	Precision	Recall	F1-Score
Equal Weight	94.93	95.48	94.79
Constraint Condition	99.09	99.85	99.47
Suspected Violation	96.24	97.08	96.76

Under Constraint Condition, the reported F1-score of 99.47% is the headline result. In the same setting, the best listed baselines are Mamba with F1 98.69, ViT-Transformer with F1 97.81, and XGBoost with F1 97.65. The paper also reports that general-purpose LLMs perform noticeably worse than tailored models on IFD and states that even the best fine-tuned GPT-4o and LLaMA-3.1-405B models remain below MaBoost’s performance in the strongest settings (Huang et al., 27 Jul 2025).

4. Behavioral findings and regulatory-forensics implications from Form 4 IFD

The benchmark paper presents IFD not only as a prediction problem but also as an empirical window into insider disclosure behavior. It reports 4,051,143 total trades and states that roughly 17.4%—21,482 transactions—are classified as filing violations. Among violations, 77% are described as oversight and 23% as intentional. The average reporting delay is 37 business days. For intentional violations, the average delay is 116 days for purchases and 77 days for sales. Average delinquent trade size is reported as \$FirmRatio = \frac{Violations\_Firm}{Trades\_Firm},$71.36 million for sales (Huang et al., 27 Jul 2025).

Role heterogeneity is explicit. The paper reports the following role-based violation rates: CEO, 10.88% with 1,397 violations over 12,843 trades; Corporate Suite, 10.63% with 2,277 violations over 21,415 trades; Beneficial Owners, 22.24% with 4,093 violations over 18,402 trades; Other Insiders, 16.59% with 1,405 violations over 8,472 trades. Beneficial owners therefore exhibit the highest listed violation rate. The paper further reports that over 56,000 trades are part of “stealth sequences,” 697 round-trip trades occur before reporting, about 43,000 trades take place within 60 days before earnings announcements but are disclosed only after the announcement, and 21 trades occur during fraudulent restatement periods with delays of hundreds of days (Huang et al., 27 Jul 2025).

Economic magnitude is part of the paper’s regulatory-forensics framing. Late-filed purchases are reported to earn 0.03% per day, with 4.64% cumulative abnormal return over the delay window. These descriptive results are used to support the claim that delayed filings are substantial in economic magnitude and often clustered around sensitive corporate events. A plausible implication is that the benchmark’s legal labels are not merely technical artifacts of filing administration; they also identify episodes with nontrivial informational and market consequences (Huang et al., 27 Jul 2025).

Interpretability is addressed through tree-based feature importance and feature-group ablation. The paper highlights gain, split counts, and SHAP values for XGBoost, and it removes groups such as Insider History, Trade Characteristics, Governance, Financial Health, and Spatiotemporal to measure performance changes. In Equal Weight, removing Insider History reduces F1 from 94.79 to 86.09, and removing Spatiotemporal reduces it to 82.33. In Constraint Condition, removing Financial Health reduces F1 from 99.47 to 98.65, while removing Governance reduces it to 87.52. In Suspected Violation, removing Insider History reduces F1 from 96.76 to 91.76, removing Trade Characteristics reduces it to 83.21, removing Governance reduces it to 74.06, and removing Spatiotemporal yields 64.79. These ablations indicate that Insider Ratio, FirmRatio, Ln(Distance), Gap_Days, governance measures, and financial health variables are central to high-performing detection in different regulatory settings (Huang et al., 27 Jul 2025).

5. Reporting inversion, predictive decoupling, and the broader IFD framework

The Form 144/Form 4 study generalizes IFD from late execution reports to a wider informational-friction regime. Form 4 is characterized as near-real-time disclosure of actual insider trades because it must be filed within two business days of execution. Form 144, by contrast, is only a notice of proposed sale, opens a 90-day statutory window, and carries no mandatory follow-up if the sale does not occur. The paper therefore defines reporting inversion as the systematic reversal in which execution is disclosed before, or essentially without, timely disclosure of intent. This generates predictive decoupling: public signals no longer permit a reliable mapping from declared intent to eventual execution (Neupane, 19 Feb 2026).

The paper places this structure inside a welfare-economic framework built around the First Fundamental Theorem of Welfare Economics and the requirement of symmetric information. Its empirical design has three principal components. First, it defines the filing gap

$FirmRatio = \frac{Violations\_Firm}{Trades\_Firm},$8

and measures cumulative market excess return during the intent-to-execution interval as

$FirmRatio = \frac{Violations\_Firm}{Trades\_Firm},$9

Over 2.6M matched filings, it reports $$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$0, $$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$1, $$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$2, indicating that insiders systematically execute during periods of positive market excess performance (Neupane, 19 Feb 2026).

Second, the paper estimates expected returns using the Carhart four-factor model over $$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$3 to $$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$4 relative to the initial Form 144 filing and compares cumulative abnormal returns around Form 144 and Form 4. It reports $$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$5, $$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$6, and

$$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$7

The interpretation given is that the market treats Form 144 as largely uninformative or routine, while the bulk of the negative adjustment occurs when Form 4 reveals actual execution. The paper terms the resulting wedge a quantified IFD effect, specifically a 1.3% “information premium” between when the market should react to intent and when it actually does (Neupane, 19 Feb 2026).

Third, the paper studies non-execution. It defines a non-execution when cumulative executed volume within the 90-day window is below proposed volume net of tolerance. Because there is no mandatory confirmation of abort, the market only learns that the intent failed when time passes and no Form 4 appears. The paper’s machine-learning audit treats non-execution as the target class and evaluates ten model families under severe class imbalance using SMOTE and cost-sensitive learning. It reports a false negative rate of approximately 46.77% for non-executions in the XGBoost confusion matrix and a broader plateau of approximately 52.4% under robustness checks, which the paper interprets as an opacity rate and a structural information ceiling. Even when balanced random forests, XGBoost, NODE, TabNet, FT-Transformer, TFT, DNDF, elastic net, weighted SVM, and Isolation Forest are used, PR-AUC and recall plateau rather than converging to reliable discrimination. In the paper’s formulation, this means the market faces genuine Knightian uncertainty about whether a filed intent will be executed, so disclosure incompleteness itself becomes a form of IFD (Neupane, 19 Feb 2026).

The broader empirical consequences extend beyond returns. Using DGTW-adjusted calendar-time portfolios, the paper reports that equal-weighted non-execution portfolios produce 32.21 bps per month with $$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$8, $$Trade\ Value = \frac{Trade\_Amount}{Market\_Cap}, \quad Ln(Distance) = \ln(\text{Distance}),$$9, while value-weighted alpha is 0.46 bps, $$Blockholder\ Ratio = \frac{Shares\_Blockholders}{Shares}, \quad HHI = \sum_{k=1}^{n} (Ownership\_Share_k)^2,$$0. In the size split, small-cap alpha is 30.96 bps, $$Blockholder\ Ratio = \frac{Shares\_Blockholders}{Shares}, \quad HHI = \sum_{k=1}^{n} (Ownership\_Share_k)^2,$$1, $$Blockholder\ Ratio = \frac{Shares\_Blockholders}{Shares}, \quad HHI = \sum_{k=1}^{n} (Ownership\_Share_k)^2,$$2, whereas large-cap alpha is 14.49 bps, $$Blockholder\ Ratio = \frac{Shares\_Blockholders}{Shares}, \quad HHI = \sum_{k=1}^{n} (Ownership\_Share_k)^2,$$3, $$Blockholder\ Ratio = \frac{Shares\_Blockholders}{Shares}, \quad HHI = \sum_{k=1}^{n} (Ownership\_Share_k)^2,$$4, which the paper terms the large-cap significance paradox. In cross-sectional regressions, prior idiosyncratic volatility carries a coefficient of −0.0034, $$Blockholder\ Ratio = \frac{Shares\_Blockholders}{Shares}, \quad HHI = \sum_{k=1}^{n} (Ownership\_Share_k)^2,$$5, and causal estimators of illiquidity effects report treatment effects around 0.0176–0.0415, with tail multipliers for large signal magnitudes up to 2.63× the mean effect. The paper uses these results to argue that IFD in the Form 144/Form 4 setting is not only a matter of delayed price discovery but also of degraded liquidity and persistent informational rents (Neupane, 19 Feb 2026).

6. Limitations, policy responses, and future research

Both papers present IFD as a compliance-relevant and research-relevant object but also delimit their scope. The Form 4 benchmark includes only U.S. public firms, open-market trades, and high-quality records, excluding options, awards, and some extreme events after cleaning and winsorization. The Form 144/Form 4 study depends on LSEG Insider, BoardEx, CRSP, and Compustat linkages, focuses on insider sales under Rule 144, and is specific to the U.S. SEC regime. Both papers therefore imply restricted external validity across jurisdictions and filing systems. They also note that even strong predictive performance does not eliminate false positives or false negatives, which is consequential in compliance settings with legal and economic stakes (Huang et al., 27 Jul 2025, Neupane, 19 Feb 2026).

A recurrent misconception is that IFD is only a timestamp problem on Form 4. The narrower benchmark does define IFD as delayed Form 4 disclosure beyond the two-business-day deadline, but the second paper argues for a broader theory in which filing order, unresolved intent, and non-confirmation are equally important. Another misconception is that general-purpose LLMs are adequate substitutes for task-specific compliance models. The benchmark results state that LLMs perform noticeably worse than tailored architectures on structured financial and relational data, and the Form 144/Form 4 audit likewise shows that even sophisticated public-information-based learners hit a ceiling when the disclosure regime omits verifiable outcome reporting. Taken together, these results suggest that some limits are algorithmic, but others are institutional (Huang et al., 27 Jul 2025, Neupane, 19 Feb 2026).

The most explicit policy proposal is Form 144-A, described as a mandatory execution confirmation. The proposal requires that, for each Form 144 filing, at or within two business days of the end of the 90-day window, insiders file a Form 144-A indicating whether the proposed shares were fully executed, partially executed, or not executed, and, if non-executed or partially executed, provide a categorization of reasons. The paper frames this as a shift from a unilateral intent regime to a bilateral accountability regime. The intended effect is to reduce informational delay, eliminate structural opacity and phantom risk, compress the reported 1.3% CAR gap, diminish post-expiration abnormal returns such as 32.21 bps per month and 14.49 bps in large caps, and lessen illiquidity distortions that can reach 2.63× the mean effect (Neupane, 19 Feb 2026).

The benchmark paper identifies additional research directions. These include extending IFD beyond U.S. Form 4 to other jurisdictions and regulatory regimes, incorporating multilingual filings, integrating unstructured disclosures such as 10-K narratives and news with document-level embeddings, moving from predictive models to causal frameworks that identify drivers of non-compliance and simulate policy interventions, and exploring more advanced sequence models and graph-based models for relations among insiders, firms, and events. Because the benchmark also provides baseline models, LLM evaluation templates, and an open repository for data and code, it establishes IFD as a reproducible benchmark for financial compliance, regulatory forensics, and interpretable time-series classification, while the Form 144/Form 4 study extends the concept into a broader theory of asymmetric resolution of insider intent (Huang et al., 27 Jul 2025).