Explainable Anomaly Detection in Hospitality Revenue Records: A Scenario-Based Modeling Study

Performance measurement Fit Jan to Jun; calibrate Jul to Aug; evaluate Sep to Dec. Block bootstrap (300 resamples of property-weeks); three replicate datasets; seven evaluation scenarios. Figure 1. Study design. Development data (fit and calibration) and evaluation data are separated in time; stress scenarios replace the evaluation period only. 2.2 Data-generating mechanism Properties and stays. Four properties with 160, 120, 90, and 200 rooms and average daily rates of 145, 110, 85, and 175 currency units were simulated for every day of 2025. Daily arrivals followed a Poisson distribution with a seasonal mean (peak in spring and summer, about 72% average occupancy) and a length of stay of one to ten nights. Each folio received a nightly room charge posted by the night audit between 02:30 and 03:10, an optional daily breakfast package at a fixed price, restaurant, bar, and room-service checks, and occasional spa, minibar, laundry, and parking charges. Bar checks were drawn from a fixed menu so that identical amounts occur naturally, and 8% of bar checks were followed by a second identical round within 2 to 25 minutes. Each property had 14 front-office and outlet clerks plus an audit user. Legitimate credits and look-alikes. About 4% of folios received an allowance against a charge; 12% of those were service-recovery comps of 50% to 100% of the charge. Allowances of 50 or more required approval, which was recorded 92% of the time, and 6% of allowances were posted between midnight and 06:00 as audit corrections. About 3% of folios received a refund, of which 6% were dispute refunds 8 to 30 days after checkout and 6% were unlinked deposit refunds. A further 0.8% of folios received an unlinked loyalty goodwill credit in a round amount. These patterns were included on purpose: each resembles one of the anomaly types and is a realistic source of false alerts. Anomalies. Anomalies were injected at a rate of 0.5% of charge records in the development period and at the scenario rate in the evaluation period. Each anomaly was attached to a randomly chosen host charge and assigned to one of three families and ten types (Table 2). For refunds and adjustments, 60% were posted by one of two clerks per property and 35% were posted between midnight and 05:00, reflecting the concentration of misconduct in a few users and outside supervised hours. Duplicates kept the clerk of the original record. Table 2. Anomaly taxonomy and counts. Family and type Definition All data Evaluati on Refund: over-refund Refund 1.2 to 2.5 times the linked charge 312 99 Refund: unlinked Refund with no matching charge 307 93 Refund: late Refund 10 to 60 days after checkout 176 59 Refund: repeat Second partial refund taking the total above the charge 253 77 Adjustment: large share Allowance of 50% to 100% of a charge, usually unapproved 427 116 Adjustment: threshold hugging Allowance of 42.00 to 49.99, just below the approval limit 369 125 Adjustment: round unlinked Credit of 50, 75, 100, 150, or 200 with no linked charge 262 78 Duplicate: exact Same charge reposted within 6 minutes 385 121 Duplicate: delayed Same charge reposted 1 to 30 hours later 405 147 Duplicate: room night Second room charge for the same night 61 22 Total 2,957 937 Evaluation counts refer to September to December 2025. Anomalies make up 0.49% of all records. 2.3 Separation of development and evaluation data Records were split by posting date. Records from January to June (289,192 records, 1,408 anomalies) were used to fit models and to compute peer-group medians and median absolute deviations. Records from July to August (118,649 records, 576 anomalies) were used only to choose thresholds. Records from September to December (189,988 records, 937 anomalies, 72 property-weeks) were used only for evaluation. Credits posted after December 31, 2025 for December stays (2,683 records) were excluded. All features were computed from information available at posting time: duplicate flags look only at earlier records, and clerk activity uses the previous 30 days. Nothing estimated on evaluation data was used by any detector. 2.4 Transparent indicators All detectors draw on the same 17 indicators, each of which a reviewer can check on the folio itself (Table 3). One candidate indicator, a flag for any repeat of the same item and amount on a folio within 36 hours, was dropped during development because it fired on 32.3% of normal records: nightly room charges and daily breakfast packages repeat by design. It was replaced by two narrower flags (repeat on the same business date; repeat of a non-recurring item). This is a first, small illustration of sensitivity to transaction patterns, and of the data-quality point that a key must reflect how records are actually produced (Alieme et al., 2022; Ilodigwe & Adesemoye, 2021). Table 3. Indicators available to all detectors. Indicator Definition Record is a refund / adjustment Transaction type flags Credit ratio Credit amount divided by the linked charge Cumulative credit ratio All credits so far against the linked charge divided by that charge Unlinked credit Credit with no linked charge Days after checkout Posting time minus checkout time, for credits Off-hours credit Credit posted between 00:00 and 06:00 Unapproved credit at or above limit Credit of 50 or more with no approval code Credit just under limit Credit between 42.50 and 49.99 Round credit Credit of 50 or more that is a multiple of 25 Exact repeat Same folio, item, and amount within 10 minutes Same-day repeat Same folio, item, and amount on the same business date, after 10 minutes Non-recurring repeat Same folio, item, and amount within 36 hours, excluding room and breakfast Room-night repeat Second room charge on a folio for the same night Clerk credit activity Clerk's share of credits over the prior 30 days, robust z within property Credits on folio Number of credits already posted to the folio Amount versus peers Robust z of amount within property, type, and item (development medians) 2.5 Detectors R1, policy rule set. A record is flagged if any of eight rules fires: unlinked credit; cumulative credit above the linked charge; credit more than seven days after checkout; adjustment of 50% or more of a charge; unapproved credit at or above the limit; exact repeat; same-day repeat; room- night repeat. R1 has no tunable threshold and stands for a typical exception report of the kind recommended in internal control and fraud risk guidance (COSO & ACFE, 2023; COSO, 2013). R2, expert scorecard. Fourteen yes or no indicators carry points fixed before any results were seen: 4 points each for an exact repeat and a second room charge for the same night; 3 each for an unlinked credit, a cumulative credit above the linked charge, and an unapproved credit at or above the limit; 2 each for a same-day repeat, a non-recurring repeat, a credit more than seven days after checkout, an adjustment of 50% or more, and an off-hours credit; and 1 each for a credit just under the limit, a round credit amount, high clerk credit activity, and two or more credits on the folio. The score is the sum, and the threshold is chosen on calibration data. The design follows the logic of points-based scorecards used in credit decisions (Siddiqi, 2006; Ustun & Rudin, 2016). S1, peer robust z-score. Each record receives the largest of four robust z-scores: amount against its peer group, cumulative credit ratio, days after checkout, and clerk credit activity. S1 needs no labels and reflects a common audit-analytics screen based on the median and median absolute deviation (Iglewicz & Hoaglin, 1993; Leys et al., 2013). S2, sparse logistic regression. An L1-penalized logistic regression (C = 0.5, balanced class weights, standardized inputs; ratios and counts log-transformed) was fitted on the January to June data with the liblinear solver (Fan et al., 2008; Friedman et al., 2010; Tibshirani, 1996). Each alert can be explained as a sum of indicator contributions, read in the usual way as odds multipliers (Hosmer et al., 2013). S3, shallow decision tree. A classification tree with maximum depth 4, at least 50 records per leaf, and balanced class weights. Each alert follows one path of at most four questions (Breiman et al., 1984). Reference, isolation forest. 200 trees fitted to a sample of 150,000 development records, without labels. It is included to show what an opaque, widely used method achieves on the same indicators (Liu et al., 2008, 2012). 2.6 Operating points and reviewer workload Two thresholds per detector were fixed on the calibration months. OP1 maximizes F1, the harmonic mean of precision and recall. OP2 is the lowest threshold whose alert rate stays within a reviewer capacity of 30 alerts per property-week. For detectors with few distinct score values (R1, R2), OP2 can coincide with OP1. Workload was translated into time with an assumed 5 minutes of triage per alert, so 30 alerts equal 2.5 hours per property-week. We report alerts per 1,000 records, alerts and hours per property-week, and alerts per true positive (the number of alerts a reviewer reads to find one real anomaly). Fixing a review budget in this way mirrors practice in card-fraud operations, where investigators can examine only a fixed number of alerts per day (Dal Pozzolo et al., 2014, 2018). 2.7 Class imbalance For S2 and S3 we compared three training strategies: no adjustment, balanced class weights, and random undersampling of normal records to 20 per anomaly (Dal Pozzolo et al., 2015; He & Garcia, 2009; King & Zeng, 2001). Each was assessed at the default probability threshold of 0.5 and at a threshold tuned on calibration data. Separately, all detectors were evaluated with thresholds held fixed while anomaly prevalence in the evaluation period was set to 0.1%, 0.5%, and 2.0%. 2.8 Sensitivity to transaction patterns Four pattern shifts were applied to the evaluation period only, leaving models and thresholds unchanged (Table 4). Table 4. Evaluation scenarios. Scenario What changes from September to December Baseline Same mechanism as development; 0.5% anomalies Low / high prevalence Anomaly rate 0.1% or 2.0% Group-booking surge On 60% of days each hotel hosts a group: 6 to 14 rooms billed nightly to one master folio at the same rate Service-recovery promotion Allowance frequency triples (4% to 12% of folios) and 60% of allowances are 30% to 60% of the charge, with approval 15% price increase All charge prices rise 15%, while peer medians remain those of January to June Concealment Anomalies shift to the subtlest variants: repeat partial refunds, threshold-hugging allowances, delayed duplicates; posted in business hours; only 40% by the two concentrated clerks 2.9 Performance measures and uncertainty Precision is TP / (TP + FP), recall is TP / (TP + FN), and the false-positive rate is FP / (FP + TN), where TP, FP, FN, and TN count records. PR-AUC is the average precision over the full score ranking. Because records within a property and week share guests, clerks, and events, 95% intervals were obtained by a block bootstrap that resamples whole property-weeks (300 resamples) (Efron & Tibshirani, 1993; Künsch, 1989; Politis & Romano, 1994). The full pipeline was repeated on two further datasets generated with different random seeds. 2.10 Implementation The generator, indicators, models, and evaluation were written in Python (NumPy, pandas, scikit- learn) (Pedregosa et al., 2011), using conventions from standard statistical learning texts (Hastie et al., 2009). Every random step is seeded, and the scripts reproduce all tables and figures in this paper. The default parameter values are those stated in Sections 2.2 to 2.9. 3. Results and Discussion 3.1 Simulated ledgers The baseline dataset contained 600,512 records: 303,836 food and beverage charges, 185,588 room charges, 104,213 other charges, 3,863 adjustments, and 3,012 refunds. The evaluation period averaged 2,639 records and 13.0 anomalies per property-week, so a perfect detector would raise 13 alerts per property-week. 3.2 Detection accuracy Table 5 reports held-out performance at OP1. The policy rule set R1 found 88% of anomalies but generated 2,821 false alerts against 823 true ones. The sparse logistic model S2 achieved the highest F1 and PR-AUC: 84% of its alerts were real, it found 72% of anomalies, and it flagged 0.07% of normal records. The scorecard R2 was the most precise detector but missed about half of all anomalies. The decision tree S3 sat between S2 and R2. The unsupervised peer screen S1 and the isolation forest both trailed the supervised transparent models, which shows that interpretability did not cost accuracy in this setting; the opaque reference was beaten by two transparent methods. Table 5. Held-out performance at OP1 (threshold maximizing F1 on calibration data), September to December 2025. Detector Precision (95% CI) Recall (95% CI) FPR F1 PR- AUC TP / FP R1 Rule set 0.226 (0.212, 0.241) 0.878 (0.862, 0.895) 1.49 % 0.35 9 0.199 823 / 2,821 R2 Scorecard 0.945 (0.923, 0.966) 0.512 (0.479, 0.545) 0.01 % 0.66 4 0.600 480 / 28 S1 Peer robust z 0.562 (0.516, 0.604) 0.409 (0.374, 0.440) 0.16 % 0.47 3 0.377 383 / 299 S2 Sparse logistic 0.839 (0.804, 0.871) 0.722 (0.686, 0.751) 0.07 % 0.77 6 0.747 676 / 130 S3 Shallow tree 0.721 (0.682, 0.756) 0.605 (0.579, 0.636) 0.12 % 0.65 8 0.634 567 / 220 Isolation forest (ref.) 0.608 (0.575, 0.643) 0.502 (0.460, 0.535) 0.16 % 0.55 0 0.537 470 / 303 FPR, false-positive rate. Intervals from 300 block-bootstrap resamples of property-weeks. 937 anomalies among 189,988 records. Figure 2. Precision-recall curves on the evaluation period. Dots mark the OP1 threshold chosen on calibration data. R1 has a single operating point. 3.3 Reviewer workload Table 6 converts each detector's alerts into review time. At OP1, S2 asked for 11.2 alerts per property-week, under one hour, and reviewers read 1.19 alerts for each real anomaly. R1 asked for 4.5 times as much review for 16 percentage points more recall. Across the four hotels over a year, that difference is roughly 877 hours of triage for R1 against 194 hours for S2. At OP2, where thresholds were relaxed to use the full 30-alert capacity on calibration data, recall rose for the statistical models (S3 to 0.802, S2 to 0.748) but precision fell by about half, and evaluation workloads of 21 to 24 alerts per property-week stayed inside capacity. Figure 3 shows the full trade-off: for any number of alerts a team can read each week, it shows the share of anomalies that would be found. S2 dominates at workloads near the true anomaly count; S3 overtakes it between roughly 17 and 60 alerts per property-week; R1 only becomes competitive above 50. Table 6. Reviewer workload on the evaluation period at both operating points. Detector Op. Prec. Recal l FPR Alerts / prop- week Hours / prop- week Alerts per TP Hours per year (4 hotels) R1 Rule set fixed 0.226 0.878 1.49 % 50.6 4.22 4.43 877 R2 Scorecard OP1 = OP2 0.945 0.512 0.01 % 7.1 0.59 1.06 122 S1 Peer robust z OP1 0.562 0.409 0.16 % 9.5 0.79 1.78 164 S1 Peer robust z OP2 0.317 0.543 0.58 % 22.3 1.86 3.15 386 S2 Sparse logistic OP1 0.839 0.722 0.07 % 11.2 0.93 1.19 194 S2 Sparse logistic OP2 0.407 0.748 0.54 % 23.9 1.99 2.46 414 S3 Shallow tree OP1 0.721 0.605 0.12 % 10.9 0.91 1.39 189 S3 Shallow tree OP2 0.497 0.802 0.40 % 21.0 1.75 2.01 364 Isolation forest OP1 0.608 0.502 0.16 % 10.7 0.89 1.64 186 Isolation forest OP2 0.320 0.595 0.63 % 24.2 2.01 3.12 419 Prec., precision. Hours assume 5 minutes of triage per alert. Annual hours = hours per property- week x 4 properties x 52 weeks. R2 has no threshold between 4 and 5 points, and a threshold of 4 exceeded capacity on calibration data, so OP2 equals OP1. Figure 3. Recall achieved for a given weekly review workload when records are reviewed in score order. Dotted lines mark the true anomaly rate (13 per property-week) and the assumed reviewer capacity (30). 3.4 Detection by anomaly type Table 7 breaks recall down by anomaly type. Unusual refunds were easy for supervised models (S2 recall 0.91 across refund types) and for R1, whose rules mirror their definitions. Adjustments were caught at 0.88 by S2, which was notably better than R1 on threshold hugging (0.93 against 0.72) because it combines the near-limit flag with clerk activity and posting time. Duplicates were the clear weak point: S2 found only 34% of them. Exact and delayed duplicates share every recorded field with legitimate repeat orders, so any model that keeps precision high must give these indicators modest weight. R1 caught every exact duplicate, but its duplicate rules alone produced 37.6 alerts per property-week at 7.8% precision (12.8 alerts per real duplicate). Room- night duplicates were caught by all supervised methods because legitimate folios never carry two room charges for one night. The type profile also exposes a blind spot in the tree. S3 never split on days after checkout, so it found only 14% of late refunds and 10% of repeat refunds. The unsupervised methods missed duplicates almost entirely, because a duplicated charge has a perfectly ordinary amount. Table 7. Recall by anomaly type at OP1 on the evaluation period. Anomaly type (n) R1 R2 S1 S2 S3 IF Refund: over-refund (99) 1.00 0.61 0.84 0.92 0.88 0.96 Refund: unlinked (93) 1.00 0.74 0.60 0.88 0.90 0.58 Refund: late (59) 1.00 0.29 1.00 0.81 0.14 0.95 Refund: repeat (77) 1.00 0.69 0.56 1.00 0.10 0.90 Adjustment: large share (116) 1.00 0.38 0.20 0.85 0.86 0.56 Adjustment: threshold hugging (125) 0.72 0.57 0.54 0.93 0.80 0.61 Adjustment: round unlinked (78) 1.00 0.92 0.64 0.83 0.92 0.71 Duplicate: exact (121) 1.00 0.32 0.01 0.32 0.32 0.00 Duplicate: delayed (147) 0.46 0.22 0.01 0.25 0.32 0.00 Duplicate: room night (22) 1.00 1.00 0.00 1.00 1.00 0.00 Family: refunds (328) 1.00 0.91 Family: adjustments (319) 0.89 0.88 Family: duplicates (290) 0.73 0.34 3.5 Model explanations Figure 4 shows the S2 weights as odds multipliers per standard deviation of each indicator. The penalty at C = 0.5 retained all 17 indicators, but four of them (posting hour, amount against peers, adjustment flag, prior credits on the folio) had multipliers between 0.93 and 1.09 and contribute little. The strongest signals were the cumulative credit ratio (3.88), an unlinked credit (2.00), and repeats of non-recurring items (1.86) or within ten minutes (1.63). Two weights are negative and need care when explaining alerts: once the cumulative ratio is known, the single-credit ratio and the refund flag act as corrections, because legitimate refunds are often large single credits that do not exceed the charge. Reviewers should read S2 explanations as a set of contributions rather than as independent rules, and a deployment could add sign constraints so that every weight agrees with audit intuition. Figure 4. Weights of the sparse logistic model (S2) expressed as odds multipliers per standard deviation. Blue indicators raise the score; red indicators lower it. The tree S3 had 12 leaves. Its first question is whether the record is a credit; non-credits are then routed by the three duplicate flags, and credits by amount against peers and cumulative credit ratio. This is easy to read but, as Table 7 showed, it ignores timing. The scorecard's threshold sweep (Table 8) shows why R2 behaved as a high-precision filter. Every score below 5 contained large numbers of legitimate repeat bar orders (4 points for an exact repeat) and goodwill credits; above 5, nearly every alert was real. A reviewer can see in one line why an alert scored 5: for example, an unapproved credit (3) posted off-hours (2). Table 8. Scorecard (R2) operating points on the evaluation period. Minimum score 1 2 3 4 5 6 7 8 Precision 0.195 0.202 0.221 0.206 0.945 0.977 0.971 1.000 Recall 0.990 0.979 0.828 0.728 0.512 0.413 0.247 0.156 Alerts per property- week 66.0 63.1 48.7 46.1 7.1 5.5 3.3 2.0 3.6 Class imbalance Table 9 shows that handling imbalance changed thresholds more than it changed ranking. At the default threshold of 0.5, balanced class weights turned S2 into a near-exhaustive screen (recall 0.977) with 65 alerts per property-week at 19.5% precision, because weighting inflates predicted probabilities. Once the threshold was tuned on calibration data, all three strategies settled at 10 to 12 alerts per property-week. Contrary to common practice, the unadjusted model ranked anomalies best (PR-AUC 0.845 against 0.747 with weighting) and reached precision 0.907 with recall 0.685 after tuning. Undersampling was close to the unadjusted model and trains on far fewer records. The pre-specified weighted model, used as S2 elsewhere in this paper, is therefore a conservative choice. Table 9. Class-imbalance strategies for the statistical models, evaluation period. Model and strategy PR- AUC Precisio n at 0.5 Recall at 0.5 Alerts / pw at 0.5 Precisio n tuned Recall tuned Alerts / pw tuned S2, no adjustment 0.845 0.888 0.705 10.3 0.907 0.685 9.8 S2, class weighting 0.747 0.195 0.977 65.1 0.839 0.721 11.2 S2, undersampling 1:20 0.834 0.454 0.870 25.0 0.818 0.755 12.0 S3, no adjustment 0.665 0.873 0.674 10.1 0.873 0.674 10.1 S3, class weighting 0.634 0.188 0.984 68.1 0.720 0.605 10.9 S3, undersampling 1:20 0.717 0.425 0.859 26.3 0.774 0.689 11.6 Tuned thresholds maximize F1 on July to August data. pw, property-week. The S2 class-weighting row comes from a separate refit and matches Table 5 to within 0.001. Prevalence had a larger effect than training strategy. With thresholds held fixed, recall barely moved as the anomaly rate changed, but precision tracked prevalence closely (Figure 5). At 0.1% prevalence, S2 precision fell to 0.57 and R1 precision to 0.06, meaning about 16 alerts for every real one; the scorecard kept precision at 0.86 but on only 1.8 alerts per property-week. At 2.0% prevalence, S2 precision rose to 0.92 and its workload to 41 alerts per property-week, above the assumed capacity. 3.7 Sensitivity to transaction patterns Figure 5 summarizes recall, precision, and workload for every detector and scenario at OP1. Three failures stand out. • Price change broke the peer screen. A 15% price increase pushed most charges beyond the January to June peer medians. S1 then flagged 14% of all records (375 alerts per property-week, precision 0.02). Detectors that use ratios and flags rather than absolute amounts were unaffected; S2 kept precision 0.81 and recall 0.71. • Group billing broke duplicate keys. Master folios that carry several rooms at the same rate on the same night look exactly like room-night and exact duplicates when the key is folio, item, and amount. Every detector that uses those flags collapsed: S2 precision fell to 0.09 with 98 alerts per property-week. Adding the room number to the duplicate key restored S2 to precision 0.85 at 9.9 alerts, and R2 to 0.93. R1 returned to 43.5 alerts per property-week, close to its usual level. • A legitimate promotion inflated adjustment alert. When approved allowances of 30% to 60% became common, the tree S3 (which leans on amount against peers) doubled its alerts to 19.7 per property-week and its precision fell to 0.37. S2 was more robust (precision 0.75 at 12.4 alerts), because approval status offsets the large ratio. Under concealment, where anomalies used only their subtlest forms, R1 recall fell from 0.88 to 0.75, S3 from 0.61 to 0.40, and R2 from 0.51 to 0.42, while S2 held at 0.71 with unchanged precision (0.85). The isolation forest was stable across most scenarios, but at a consistently lower level than S2. Figure 5. Recall, precision, and alerts per property-week for each detector (columns) under each evaluation scenario , with thresholds fixed from calibration (OP1). IF, isolation forest. 3.8 Replication Repeating the full pipeline on two more datasets gave the same ordering (Table 10). S2 precision ranged from 0.836 to 0.861 and recall from 0.666 to 0.738 across the three datasets. The tree was the least stable method, with recall from 0.605 to 0.741 and workload from 10.9 to 17.0 alerts per property-week, a familiar property of single trees whose top splits change with small changes in data. Table 10. Performance at OP1 across three independently generated datasets: mean (minimum, maximum). Detector Precision Recall PR- AUC Alerts / property- week R1 Rule set 0.225 (0.225, 0.226) 0.885 (0.878, 0.888) 0.200 51.8 (50.6, 52.9) R2 Scorecard 0.934 (0.923, 0.945) 0.512 (0.504, 0.520) 0.594 7.2 (7.1, 7.5) S1 Peer robust z 0.547 (0.535, 0.562) 0.465 (0.409, 0.519) 0.411 11.3 (9.5, 13.0) S2 Sparse logistic 0.845 (0.836, 0.861) 0.708 (0.666, 0.738) 0.748 11.1 (10.2, 11.9) S3 Shallow tree 0.680 (0.575, 0.743) 0.655 (0.605, 0.741) 0.618 13.0 (10.9, 17.0) Isolation forest 0.660 (0.608, 0.687) 0.496 (0.487, 0.502) 0.547 10.0 (9.4, 10.7) 3.9 Two-tier review queue The results suggest pairing the scorecard's precision with the logistic model's coverage. In the recommended design, Tier 1 contains records with a scorecard score of 5 or more, each delivered with the list of indicators that earned its points. Tier 2 contains the remaining S2 alerts at OP1, each delivered with its three largest indicator contributions. Both thresholds were fixed before evaluation, so Table 11 is a held-out result for the combined model. Table 11. Evaluated two-tier model and its components, evaluation period. Queue Precisi on Recall FPR Alerts / prop- week Hours / prop- week Alerts per TP Tier 1: scorecard score 5 or more 0.945 0.512 0.01% 7.1 0.59 1.06 Tier 2: added S2 alerts (OP1) 0.658 0.226 0.06% 4.5 0.37 1.52 Combined two-tier model 0.834 0.739 0.07% 11.5 0.96 1.20 For comparison: R1 refund and adjustment rules only 0.655 0.653 0.17% 13.0 1.08 1.53 For comparison: R1 duplicate rules only 0.078 0.225 1.32% 37.6 3.14 12.84 Recall of Tier 2 is the additional share of all anomalies found beyond Tier 1. The combined model found 74% of anomalies with 83% of alerts real, at 11.5 alerts and just under one hour per property-week, about 200 hours a year across the four hotels. It matched S2 alone on workload while ordering the work so that the most certain cases come first. Its remaining gap is duplicates, which Section 3.11 discusses. 3.10 Principal findings In this simulation, transparent methods identified unusual refunds and adjustments well and at modest workloads. A sparse logistic model built on 17 checkable indicators found about 72% of all anomalies, and about 90% of refund and adjustment anomalies, while asking a reviewer to read roughly 11 records per hotel per week. An expert scorecard identified half of all anomalies with almost no false alerts. A two-tier queue that combines them reached 74% recall at 83% precision for under an hour of review per property-week. A conventional exception report achieved higher recall but at four to five times the workload, mostly through duplicate rules that flag legitimate repeat orders. How well transparent detection works depends on the anomaly family. For refunds and adjustments, the information needed to separate legitimate from irregular activity is present in the ledger: link to an original charge, cumulative amount, timing, approval, and who posted it. For duplicates, the ledger fields used here are not enough, and better models do not fix missing information. The opaque isolation forest ranked below two transparent supervised methods on every headline measure. This supports the view that, for structured records with meaningful features, interpretable models can perform as well as complex ones (Breiman, 2001; Caruana et al., 2015; Rudin, 2019). This is also what audit guidance for AI-supported financial systems asks for: evidence that can be validated and human oversight of automated judgments (Elebe & Anunagba, 2025; Fobellah, 2025b; Ominyi & Anichukwueze, 2023). The benefit of transparency here is practical: every alert in the recommended model arrives with reasons a reviewer can check against the folio in a few minutes, and the same reasons can be shared with the employee involved. 3.11 Duplicate entries and data keys Two results point the same way. Duplicate detection was poor when legitimate repeats shared all key fields, and it collapsed when group billing created more such repeats; adding the room number to the key fixed the second problem completely. In practice, the point-of-sale check number, the terminal, and the original posting reference would distinguish a reposted check from a second round of drinks. Hotels considering duplicate analytics should first confirm that these identifiers flow into the ledger extract, which is a data integration and governance task as much as an analytic one (Aliliele et al., 2025b; Badmus et al., 2019; Mbonu et al., 2021a). Without them, duplicate alerts should go to a separate low-priority queue so that they do not crowd out credit alerts. 3.12 Thresholds and changing transaction patterns The imbalance experiment showed that class weighting mostly changes where a threshold falls, and that threshold choice, not training strategy, determines workload. The prevalence experiment showed that a threshold that is right at 0.5% anomalies gives far more false alerts per real one at 0.1%. Thresholds should therefore be owned by the finance team, documented, and reviewed on a schedule, in the same way risk-based internal audit treats other control parameters (Akomolafe & Agu, 2018; Akomolafe et al., 2025; Ominyi & Anichukwueze, 2024), and in line with guidance on recalibration and drift monitoring (Gama et al., 2014; Van Calster et al., 2019), using recent labeled outcomes from the review queue. A simple control is to track weekly alerts and confirmed findings and to reset the threshold when either drifts beyond an agreed band. Each failure in Section 3.7 traces to a known business event: a rate increase, a group contract, a service-recovery campaign. These events are announced in advance, so detectors can be stress- tested before they occur, as this study did. Ratio-based indicators survived the price change while absolute peer baselines did not, which argues for rebasing peer statistics after any price change and for preferring relative indicators where possible. This is the dataset-shift problem described in the machine learning literature, here made concrete with hotel events (Gama et al., 2014; Moreno- Torres et al., 2012). 3.13 Implications for practice and research Earlier work in this line of inquiry assessed revenue-integrity risk conceptually: which points in the hotel revenue cycle are exposed, which controls are expected, and which red flags auditors should watch. That work sits within a wider body of conceptual models for technology-enabled risk assessment, audit analytics, and transaction anomaly detection (Abetoh & Atakpa, 2024; Akomolafe & Agu, 2018; Atakpa & Fobellah, 2023; Ominyi et al., 2024), most of which propose architectures without measuring how they perform. This study moves that assessment to testable transaction analytics. Each conceptual red flag became a computable indicator; each control expectation became a rule or a weight; and each claim about detection became a measured precision, recall, false-positive rate, and workload on data that the methods had not seen. The scenario generator also turns qualitative concerns, such as group billing or promotions, into tests that can be repeated whenever a property changes how it operates. The natural next step is a pilot on a hotel's own ledger extract, using the two-tier model as a starting point and the review outcomes as labels. 3.14 Limitations and future research The data are synthetic. The anomaly definitions, the indicators, and the policy rules were written by the same authors, which favors the rule set and the supervised models; R1's perfect recall on most refund types reflects this. Absolute numbers should not be read as forecasts for a real hotel. The relative findings (precision of scorecards, the duplicate problem, threshold sensitivity, the failure modes under pattern shift) are more likely to transfer, because they arise from the structure of ledger data rather than from tuning. This is the usual standard for inference from simulation (Morris et al., 2019; Sargent, 2013). The supervised models assume labeled anomalies exist for training. Real hotels have few confirmed cases, and labels from past reviews are biased toward what earlier rules flagged. In practice the scorecard and peer screen would be used first, and a logistic model would be fitted once enough reviewed alerts had accumulated, in the human-in-the- loop manner described for other review workflows (Ladapo et al., 2022, 2024). Card-fraud studies show that such feedback loops also bias future labels, which must be managed (Dal Pozzolo et al., 2018). The workload model assumes 5 minutes per alert for every alert. Real triage time varies by alert type, and confirmed anomalies need further investigation that is not counted here. The capacity of 30 alerts per property-week is an assumption that each hotel should replace with its own figure. Only one penalty setting and one tree depth were examined, and the scorecard points were fixed by judgment rather than optimized. Other transparent models, such as generalized additive models or rule lists, might perform better (Caruana et al., 2015; Ustun & Rudin, 2016). Finally, the study treats each record independently, whereas card-fraud work shows that aggregating transactions over time and across entities can add signal; schemes that spread across many small transactions or several clerks may need folio-level or user-level aggregation. 4. Conclusion Transparent methods can find most unusual refunds and adjustments in hotel revenue records with review burdens a finance team can sustain. In a controlled simulation withheld-out evaluation, a sparse logistic model on 17 checkable indicators reached 84% precision and 72% recall at about 11 alerts and under one hour of review per property-week, and a two-tier queue combining it with an expert scorecard reached 83% precision and 74% recall at similar cost. Exception-style rules found more anomalies but quadrupled the workload. Duplicate entries remain hard to separate from legitimate repeats without point-of-sale identifiers. Thresholds must be managed as prevalence changes, peer baselines must be refreshed after price changes, and duplicate keys must reflect how a property bills group. With those safeguards, explainable detection can support routine, testable revenue assurance in hotel operations. References Abdallah, A., Maarof, M. A., & Zainal, A. (2016). Fraud detection system: A survey. Journal of Network and Computer Applications, 68, 90 to 113. Abetoh, N. F., & Atakpa, M. I. (2024). Audit analytics in healthcare financial oversight: Leveraging data science to strengthen accountability in multilateral grant ecosystems. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 10(6), 2710 to 2747. https://doi.org/10.32628/CSEIT2410791 Adelanwa, A., Basnet, A., & Anene, U. N. (2023). Predictive analytics models for financial risk detection and fraud prevention in public systems. International Journal of Advanced Multidisciplinary Research and Studies, 3(6). https://doi.org/10.62225/2583049X.2023.3.6.5969 Adesuyi, M. O., Akomolafe, O., Olaogun, B. O., Ndukwe, V. U., & Sakyi, J. K. (2024). AI-driven risk scoring model for global cross-border trade payment transactions. International Journal of Advanced Multidisciplinary Research and Studies, 4(1), 1569 to 1581. https://doi.org/10.62225/2583049X.2024.4.1.5281 Adesuyi, M. O., Akomolafe, O., Olaogun, B. O., Ndukwe, V. U., & Sakyi, J. K. (2025). AI-enabled fraud detection ecosystem model for securing international payment channels. International Journal of Advanced Multidisciplinary Research and Studies, 5(6), 866 to 882. https://doi.org/10.62225/2583049X.2025.5.6.5284 Adesuyi, M. O., Kalu, A., & Walawalkar, G. (2023). Data-led cost governance in technology- intensive enterprises. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 9(3), 877 to 896. https://doi.org/10.32628/IJSRCSEIT Adesuyi, M. O., Walawalkar, G., & Kalu, A. (2022). Predictive budgeting models using operational and market signals. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 8(5), 818 to 837. https://doi.org/10.32628/IJSRCSEIT Agboada, P., Ezeako, R., Okoli, I. E., & Uzondu, C. S. (2018). Digital transformation in operations management: A review of advances in automation, analytics, and process redesign. International Journal of Engineering Technology Research & Management, 2(11), 109 to 134. Agboada, P., Ezeako, R., Okoli, I. E., & Uzondu, C. S. (2019). Standard operating procedures as change infrastructure: A conceptual framework for sustaining operational improvement. Iconic Research and Engineering Journals, 3(6), 706 to 730. https://doi.org/10.64388/IREV3I6-1723313 Agu, M. U., Akomolafe, O., & Bello, A. (2023). A comparative review of SOX compliance frameworks in cross-border financial auditing. International Journal of Advanced Multidisciplinary Research and Studies, 3(6), 2297 to 2306. https://doi.org/10.62225/2583049X.2023.3.6.5362 Agu, M. U., Akomolafe, O., & Bello, A. (2025). Advances in predictive financial risk assessment using Python-based forecasting systems. International Journal of Computer Science and Mathematical Theory, 11(11), 131 to 148. https://doi.org/10.56201/ijcsmt..pg131.148 Akhigbe, R., Falemi, A., & Akin-Oluyomi, O. T. (2023). A conceptual model for improving customer experience using workforce behavior and real-time operational data. International Journal of Multidisciplinary Research and Growth Evaluation, 4(6), 1322 to 1338. https://doi.org/10.54660/.ijmrge.2023.4.6.1322-1338 Akin-Oluyomi, O. T., & Akhigbe, R. (2022). A supply chain governance model for enhancing compliance and operational quality across retail networks. International Journal of Multidisciplinary Research and Growth Evaluation, 3(6), 860 to 878. https://doi.org/10.54660/.ijmrge.2022.3.6.860-878 Akin-Oluyomi, O. T., & Akhigbe, R. (2023). An advanced framework for improving multi-site operational efficiency using data-driven performance indicators. International Journal of Advanced Multidisciplinary Research and Studies, 3(1), 1763 to 1781. https://doi.org/10.62225/2583049X.2023.3.1.5343 Akinleye, O. K., & Adeyoyin, O. (2021). Process automation framework for enhancing procurement efficiency and transparency. Shodhshauryam, International Scientific Refereed Research Journal, 4(4), 356 to 387. Akomolafe, O., & Agu, M. U. (2018). A conceptual model for enhancing internal audit quality through technology-enabled risk assessment frameworks. Iconic Research and Engineering Journals, 1(9), 458 to 475. Akomolafe, O., & Agu, M. U. (2019a). A conceptual framework for developing risk-based internal control models in the insurance and banking sectors. Iconic Research and Engineering Journals, 2(8), 335 to 354. Akomolafe, O., & Agu, M. U. (2019b). A review of data-driven risk evaluation models for emerging market financial institutions. Iconic Research and Engineering Journals, 3(6), 433 to 448. Akomolafe, O., & Agu, M. U. (2019c). Advances in financial resilience through integrated governance and compliance strategies. Iconic Research and Engineering Journals, 2(10), 607 to 620. Akomolafe, O., Agu, M. U., & Bello, A. (2023a). A conceptual model for implementing risk-based auditing in strategic financial management. International Journal of Advanced Multidisciplinary Research and Studies, 3(6), 2274 to 2286. https://doi.org/10.62225/2583049X.2023.3.6.5359 Akomolafe, O., Agu, M. U., & Bello, A. (2023b). A quantitative conceptual model for assessing compliance risk in emerging economies. International Journal of Advanced Multidisciplinary Research and Studies, 3(6), 2287 to 2296. https://doi.org/10.62225/2583049X.2023.3.6.5360 Akomolafe, O., Agu, M. U., & Bello, A. (2025). A conceptual model for advancing risk governance through data-driven compliance analytics in financial institutions. Journal of Accounting and Financial Management, 11(11), 211 to 228. https://doi.org/10.56201/jafm..pg211.228 Alieme, J. I., Sule, A., & Obukadata, V. O. (2022). Standardization of reporting and decision workflows: A framework for improving efficiency, data quality, and management visibility. International Journal of Multidisciplinary Research and Growth Evaluation, 3(6), 1190 to 1218. https://doi.org/10.54660/.IJMRGE.2022.3.6.1190-1218 Aliliele, C., Eboh, E. E., & Oluwo, K. (2024b). Advances in variance analytics for cost control in manufacturing and industrial enterprises. International Journal of Advanced Multidisciplinary Research and Studies, 4(6), 3163 to 3185. https://doi.org/10.62225/2583049X.2024.4.6.6012 Aliliele, C., Mbonu, I. S., & Iwuanyanwu, U. (2023). A review of API governance and risk prioritization frameworks in modern financial institutions. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 9(10), 395 to 433. https://doi.org/10.32628/CSEIT2361072 Aliliele, C., Mbonu, I. S., & Iwuanyanwu, U. (2024a). A conceptual framework for enterprise data sensitivity classification and regulatory traceability mechanisms. International Journal of Advanced Multidisciplinary Research and Studies, 4(6), 3103 to 3124. https://doi.org/10.62225/2583049X.2024.4.6.5991 Aliliele, C., Mbonu, I. S., Uzoka, E., & Iwuanyanwu, U. (2025a). A review of AI assisted continuous auditing systems in technology risk and cybersecurity oversight. Gyanshauryam, International Scientific Refereed Research Journal, 8(4), 210 to 250. https://doi.org/10.32628/GISRRJ258369 Aliliele, C., Mbonu, I. S., Uzoka, E., & Iwuanyanwu, U. (2025b). Advances in data lakehouse governance architectures for enterprise data loss prevention and compliance assurance. Shodhshauryam, International Scientific Refereed Research Journal, 8(4), 171 to 213. https://doi.org/10.32628/SHISRRJ258474 Alles, M. G., Kogan, A., & Vasarhelyi, M. A. (2008). Putting continuous auditing theory into practice: Lessons from two pilot implementations. Journal of Information Systems, 22(2), 195 to 214. Ambali, K. B., Eyetsemitan, R. A., Oyeleye, A. O., & Fadayomi, O. (2021). Lean Six Sigma for small enterprises: A systematic review and Lite-DMAIC adaptation framework for resource-constrained organizations. Iconic Research and Engineering Journals, 5(5), 562 to 583. https://doi.org/10.64388/IREV5I5-1716957 Anichukwueze, C. C., Osuji, V. C., & Oguntegbe, E. E. (2021). Blockchain-based architectures for tamper-proof regulatory recordkeeping and real-time audit readiness. International Journal of Multidisciplinary Research and Growth Evaluation, 2(6), 485 to 504. Annan, A. O. (2024). Algorithmic accountability and trade secret protection in artificial intelligence. Shodhshauryam, International Scientific Refereed Research Journal, 7(5), 315 to 347. Annan, A. O. (2025). Automated decision-making and anti-discrimination compliance under U.S. law. International Journal of Advanced Multidisciplinary Research and Studies, 5(2), 2522 to 2540. Antonio, N., de Almeida, A., & Nunes, L. (2017). Predicting hotel booking cancellation to decrease uncertainty and increase revenue. Tourism & Management Studies, 13(2), 25 to 39. Antonio, N., de Almeida, A., & Nunes, L. (2019). Hotel booking demand datasets. Data in Brief, 22, 41 to 49. Appelbaum, D., Kogan, A., & Vasarhelyi, M. A. (2017). Big data and analytics in the modern audit engagement: Research needs. Auditing: A Journal of Practice & Theory, 36(4), 1 to 27. Association of Certified Fraud Examiners. (2024). Occupational fraud 2024: A report to the nations. ACFE. Atakpa, M. I. (2021). A review of predictive analytics models for anti-money laundering detection in commercial banking systems. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 7(2), 778 to 804. https://doi.org/10.32628/CSEIT2064813 Atakpa, M. I., & Abetoh, N. F. (2022). A systematic review of machine learning advances in financial fraud detection for banking systems. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 8(2), 771 to 801. https://doi.org/10.32628/CSEIT23906220 Atakpa, M. I., & Abolaji, T. O. (2022). A privacy-preserving data architecture model for regulated industry analytics under GDPR and HIPAA compliance. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 8(3), 781 to 808. https://doi.org/10.32628/CSEIT22558 Atakpa, M. I., & Fobellah, A. N. (2023). Anomaly detection in financial time-series data: A conceptual model for healthcare and banking applications. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 9(2), 967 to 997. https://doi.org/10.32628/CSEIT2342441 Atakpa, M. I., Abetoh, N. F., & Akeju, B. (2024). Advances in artificial intelligence for healthcare payment fraud detection: A review of NHS applications. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 10(4), 1161 to 1194. https://doi.org/10.32628/CSEIT26123240 Babatope, O. M., Akokodaripon, D. A., Akinleye, O. K., & Okoruwa, P. O. (2023). The role of data visualization in enhancing strategic procurement decision-making effectiveness. International Journal of Advanced Multidisciplinary Research and Studies, 3(6), 2205 to 2212. Badmus, O., Dosunmu, A. A., & Ozowara, D. E. (2019). A conceptual model for ETL design and data integration in Salesforce-centric enterprise architectures. Iconic Research and Engineering Journals, 3(5). Baesens, B., Van Vlasselaer, V., & Verbeke, W. (2015). Fraud analytics using descriptive, predictive, and social network techniques: A guide to data science for fraud detection. Wiley. Bao, Y., Ke, B., Li, B., Yu, Y. J., & Zhang, J. (2020). Detecting accounting fraud in publicly traded U.S. firms using a machine learning approach. Journal of Accounting Research, 58(1), 199 to 235. Barredo Arrieta, A., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., García, S., Gil-López, S., Molina, D., Benjamins, R., Chatila, R., & Herrera, F. (2020). Explainable artificial intelligence : Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion, 58, 82 to 115. Bello, A. A., Oduro, D. A., Manu, E. O., Bello, A. D., Leo, A. O., Ukatu, C. E., & Okika, N. (2025c). Enhancing Know Your Customer and Anti-Money Laundering compliance using blockchain: A business analysis approach. Iconic Research and Engineering Journals, 8(9), 297 to 305. Bello, A. D., Elebe, O., Hammed, N. I., Okoruwa, P. O., Fadayomi, O., & Omoegun, G. O. (2025b). An advanced regulatory technology framework for improving financial transparency and fraud reporting accuracy. International Journal of Advanced Multidisciplinary Research and Studies, 5(6). Bello, A. D., Oguntola, O. B., Ajibade, A. T., Akindolani, A., Ayoola, O., & Bello, A. M. (2025a). AI-driven fraud detection in UK digital payment systems: Challenges and solutions. Iconic Research and Engineering Journals, 9(1), 1329 to 1343. Bhattacharyya, S., Jha, S., Tharakunnel, K., & Westland, J. C. (2011). Data mining for credit card fraud: A comparative study. Decision Support Systems, 50(3), 602 to 613. Bola-Sadipe, D., Aliliele, C., & Eboh, E. E. (2019). Conceptual model for strategic accounting systems strengthening financial transparency and regulatory compliance. IRE Journals, 3(6), 535 to 564. https://doi.org/10.64388/IREV3I6-1715327 Bola-Sadipe, D., Aliliele, C., & Odihi, F. (2023). Conceptual model for automated financial consolidation and transparency in multinational corporations. Gyanshauryam, International Scientific Refereed Research Journal, 6(6), 414 to 455. https://doi.org/10.32628/GISRRJ236645 Bolton, R. J., & Hand, D. J. (2002). Statistical fraud detection: A review. Statistical Science, 17(3), 235 to 255. Breiman, L. (2001). Statistical modeling: The two cultures. Statistical Science, 16(3), 199 to 231. Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification and regression trees. Wadsworth. Burton, A., Altman, D. G., Royston, P., & Holder, R. L. (2006). The design of simulation studies in medical statistics. Statistics in Medicine, 25(24), 4279 to 4292. Bussmann, N., Giudici, P., Marinelli, D., & Papenbrock, J. (2021). Explainable machine learning in credit risk management. Computational Economics, 57(1), 203 to 216. Campos, G. O., Zimek, A., Sander, J., Campello, R. J. G. B., Micenková, B., Schubert, E., Assent, I., & Houle, M. E. (2016). On the evaluation of unsupervised outlier detection: Measures, datasets, and an empirical study. Data Mining and Knowledge Discovery, 30(4), 891 to 927. Carcillo, F., Le Borgne, Y.-A., Caelen, O., Kessaci, Y., Oblé, F., & Bontempi, G. (2021). Combining unsupervised and supervised learning in credit card fraud detection. Information Sciences, 557, 317 to 331. Caruana, R., Lou, Y., Gehrke, J., Koch, P., Sturm, M., & Elhadad, N. (2015). Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. In Proceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 1721 to 1730). Chandola, V., Banerjee, A., & Kumar, V. (2009). Anomaly detection: A survey. ACM Computing Surveys, 41(3), Article 15. Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16, 321 to 357. Committee of Sponsoring Organizations of the Treadway Commission, & Association of Certified Fraud Examiners. (2023). Fraud risk management guide (2nd ed.). COSO. Committee of Sponsoring Organizations of the Treadway Commission. (2013). Internal control: Integrated framework. COSO. Cressey, D. R. (1953). Other people's money: A study in the social psychology of embezzlement. Free Press. Cross, R. G., Higbie, J. A., & Cross, D. Q. (2009). Revenue management's renaissance: A rebirth of the art and science of profitable revenue generation. Cornell Hospitality Quarterly, 50(1), 56 to 81. Dal Pozzolo, A., Boracchi, G., Caelen, O., Alippi, C., & Bontempi, G. (2018). Credit card fraud detection: A realistic modeling and a novel learning strategy. IEEE Transactions on Neural Networks and Learning Systems, 29(8), 3784 to 3797. Dal Pozzolo, A., Caelen, O., Johnson, R. A., & Bontempi, G. (2015). Calibrating probability with undersampling for unbalanced classification. In 2015 IEEE Symposium Series on Computational Intelligence (pp. 159 to 166). IEEE. Dal Pozzolo, A., Caelen, O., Le Borgne, Y.-A., Waterschoot, S., & Bontempi, G. (2014). Learned lessons in credit card fraud detection from a practitioner perspective. Expert Systems with Applications, 41(10), 4915 to 4928. Davis, J., & Goadrich, M. (2006). The relationship between precision-recall and ROC curves. In Proceedings of the 23rd International Conference on Machine Learning (pp. 233 to 240). Debreceny, R. S., & Gray, G. L. (2010). Data mining journal entries for fraud detection: An exploratory study. International Journal of Accounting Information Systems, 11(3), 157 to 181. Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114 to 126. Doshi-Velez, F., & Kim, B. (2017). Towards a rigorous science of interpretable machine learning (arXiv:1702.08608). arXiv. Ebhojie, O., Dogbatsey, E. A., & Oyeleye, A. O. (2023). Audit liaison, corrective action planning, and control compliance in multinational organisations: A systematic literature review. International Journal of Advanced Multidisciplinary Research and Studies, 3(6), 2839 to 2850. https://doi.org/10.62225/2583049X.2023.3.6.6200 Eboh, E. E., & Aliliele, C. (2023). Portfolio risk optimization models for venture capital investment strategy and fraud exposure analysis. Iconic Research and Engineering Journals, 7(4), 759 to 792. https://doi.org/10.64388/IREV7I4-1715305 Eboh, E. E., & Aliliele, C. (2024). AI-driven data analytics framework for risk assessment and detection of venture capital and private equity investment fraud in U.S. capital markets. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 10(6), 2624 to 2665. https://doi.org/10.32628/CSEIT2410787 Eboh, E. E., & Aliliele, C. (2025). Predictive analytics systems for investment risk monitoring in SME FinTech companies and cross-border capital markets. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 11(2), 3969 to 4010. https://doi.org/10.32628/CSEIT25113398 Eboh, E. E., Aliliele, C., & Odihi, F. (2025). Advances in financial governance and accountability systems in small and medium enterprises. International Journal of Advanced Multidisciplinary Research and Studies, 5(6), 2223 to 2245. https://doi.org/10.62225/2583049X.2025.5.6.6013 Edivri, J., & Oteri, O. (2024). A conceptual KPI-driven decision and optimization framework for IT service delivery, portfolio performance, and adoption. Gyanshauryam, International Scientific Refereed Research Journal, 7(1), 167 to 187. https://doi.org/10.32628/GISRRJ246643 Efron, B., & Tibshirani, R. J. (1993). An introduction to the bootstrap. Chapman & Hall. Ekechi, N. V., Anunagba, C. O., & Ozowara, D. E. (2022). Advances in financial forensics techniques for detecting cyber-enabled fraud in telemedicine services. Shodhshauryam, International Scientific Refereed Research Journal, 5(1), 428 to 450. Elebe, O., & Anunagba, C. O. (2025). Responsible artificial intelligence in IT auditing through evidence validation, human oversight, and regulatory compliance. International Journal of Technology, Health and Sustainability, 1(2), 310 to 328. Elkan, C. (2001). The foundations of cost-sensitive learning. In Proceedings of the 17th International Joint Conference on Artificial Intelligence (pp. 973 to 978). Enshara, S. I., & Bastian, I. (2018). Rasionalisasi kecurangan penerimaan kas (studi kasus pada Hotel Novotel Yogyakarta) [Rationalization of cash receipt fraud: A case study at Novotel Hotel Yogyakarta]. ABIS: Accounting and Business Information Systems Journal, 6(3). https://doi.org/10.22146/abis.v6i3.59089 Ezeako, R., Ibeh, A. V., Alugwe, S., & Awulu, I. (2024). Performance dashboards and data-driven decision-making in operations management: A conceptual framework for leadership visibility. International Journal of Advanced Multidisciplinary Research and Studies, 4(6), 3520 to 3538. https://doi.org/10.62225/2583049X.2024.4.6.6930 Fadayomi, O., Bello, A. D., Elebe, O., Hammed, N. I., & Omoegun, G. O. (2024). An adaptive fraud risk scoring model for real-time transaction monitoring at scale. IIARD International Journal of Banking and Finance Research, 10(10), 212 to 230. Fan, R.-E., Chang, K.-W., Hsieh, C.-J., Wang, X.-R., & Lin, C.-J. (2008). LIBLINEAR: A library for large linear classification. Journal of Machine Learning Research, 9, 1871 to 1874. Fawcett, T. (2006). An introduction to ROC analysis. Pattern Recognition Letters, 27(8), 861 to 874. Filani, O. M., Nnabueze, S. B., Sakyi, J. K., & Okojie, J. S. (2023). Scenario-based financial modelling for enhancing strategic decision-making and organizational long-term planning. Journal of Frontiers in Multidisciplinary Research, 4(2), 251 to 265. https://doi.org/10.54660/.JFMR.2023.4.2.251-265 Fobellah, A. N. (2025a). Cognitive biases in financial decision-making: Implications for audit and risk management in large corporations: A conceptual review. International Journal of Science and Research Archive, 16(2), 17 to 22. https://doi.org/10.30574/ijsra.2025.16.2.2273 Fobellah, A. N. (2025b). Navigating digital transformation: Auditing artificial intelligence- powered financial systems: A conceptual review. International Journal of Science and Research Archive, 16(2), 23 to 28. https://doi.org/10.30574/ijsra.2025.16.2.2274 Fobellah, A. N. (2025c). Unintended consequences of financial regulations: A study of corporate compliance burdens: A conceptual review. International Journal of Science and Research Archive, 16(2), 29 to 34. https://doi.org/10.30574/ijsra.2025.16.2.2275 Friedman, J., Hastie, T., & Tibshirani, R. (2010). Regularization paths for generalized linear models via coordinate descent. Journal of Statistical Software, 33(1), 1 to 22. Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., & Bouchachia, A. (2014). A survey on concept drift adaptation. ACM Computing Surveys, 46(4), Article 44. Ghiselli, R. F., & Ismail, J. A. (1995). Gauging employee theft and other unacceptable behaviors in food service operations. Hospitality Review, 13(2), Article 3. Goldstein, M., & Uchida, S. (2016). A comparative evaluation of unsupervised anomaly detection algorithms for multivariate data. PLOS ONE, 11(4), e0152173. Guidotti, R., Monreale, A., Ruggieri, S., Turini, F., Giannotti, F., & Pedreschi, D. (2018). A survey of methods for explaining black box models. ACM Computing Surveys, 51(5), Article 93. Haixiang, G., Yijing, L., Shang, J., Mingyun, G., Yuanyue, H., & Bing, G. (2017). Learning from class-imbalanced data: Review of methods and applications. Expert Systems with Applications, 73, 220 to 239. Ham, S., Kim, W. G., & Jeong, S. (2005). Effect of information technology on performance in upscale hotels. International Journal of Hospitality Management, 24(2), 281 to 294. Hastie, T., Tibshirani, R., & Friedman, J. (2009). The elements of statistical learning: Data mining, inference, and prediction (2nd ed.). Springer. He, H., & Garcia, E. A. (2009). Learning from imbalanced data. IEEE Transactions on Knowledge and Data Engineering, 21(9), 1263 to 1284. Hilal, W., Gadsden, S. A., & Yawney, J. (2022). Financial fraud: A review of anomaly detection techniques and recent advances. Expert Systems with Applications, 193, Article 116429. Hosmer, D. W., Lemeshow, S., & Sturdivant, R. X. (2013). Applied logistic regression (3rd ed.). Wiley. Hospitality Financial and Technology Professionals. (2024). Uniform system of accounts for the lodging industry (12th rev. ed.). HFTP. Iglewicz, B., & Hoaglin, D. C. (1993). How to detect and handle outliers. ASQC Quality Press. Ilodigwe, L., & Adesemoye, A. C. (2021). The data backbone of health system transformation: A governance and architecture framework for interoperability, data quality, and advanced analytics at national scale. International Journal of Health and Pharmaceutical Research, 6(2), 52 to 76. https://doi.org/10.56201/ijhpr..pg52.76 Issa, H., Sun, T., & Vasarhelyi, M. A. (2016). Research ideas for artificial intelligence in auditing: The formalization of audit and workforce supplementation. Journal of Emerging Technologies in Accounting, 13(2), 1 to 20. Ivanov, S., & Zhechev, V. (2012). Hotel revenue management: A critical literature review. Tourism, 60(2), 175 to 197. Julianty, F., Rapina, Setiawan, S., & Ndaru, H. P. S. (2022). Fraud in accounting: Surveys on hotels in Bandung. Jurnal Riset Akuntansi Kontemporer, 14(1), 75 to 82. https://doi.org/10.23969/jrak.v14i1.4932 Kaufman, S., Rosset, S., Perlich, C., & Stitelman, O. (2012). Leakage in data mining: Formulation, detection, and avoidance. ACM Transactions on Knowledge Discovery from Data, 6(4), Article 15. Kimes, S. E. (1989). Yield management: A tool for capacity-constrained service firms. Journal of Operations Management, 8(4), 348 to 363. Kimes, S. E. (2008). The role of technology in restaurant revenue management. Cornell Hospitality Quarterly, 49(3), 297 to 309. King, G., & Zeng, L. (2001). Logistic regression in rare events data. Political Analysis, 9(2), 137 to 163. Kogan, A., Alles, M. G., Vasarhelyi, M. A., & Wu, J. (2014). Design and evaluation of a continuous data level auditing system. Auditing: A Journal of Practice & Theory, 33(4), 221 to 245. Kokina, J., & Davenport, T. H. (2017). The emergence of artificial intelligence: How automation is changing auditing. Journal of Emerging Technologies in Accounting, 14(1), 115 to 122. Komi, N. M., & Adamolekun, A. (2021). Interpretable machine learning for early failure prediction in distributed renewable energy assets. International Journal of Multidisciplinary Research and Growth Evaluation, 2(6), 1015 to 1038. https://doi.org/10.54660/.IJMRGE.2021.2.6.1015-1038 Krawczyk, B. (2016). Learning from imbalanced data: Open challenges and future directions. Progress in Artificial Intelligence, 5(4), 221 to 232. Künsch, H. R. (1989). The jackknife and the bootstrap for general stationary observations. The Annals of Statistics, 17(3), 1217 to 1241. Ladapo, O. O., Dosunmu, A. A., Jooda, D., & Abolaji, T. O. (2022). Human-in-the-loop machine learning: A state of the art. Journal of Frontiers in Multidisciplinary Research, 3(1), 656 to 669. https://doi.org/10.54660/.JFMR.2022.3.1.656-669 Ladapo, O. O., Jooda, D., Dosunmu, A. A., & Abolaji, T. O. (2024). Keeping humans in the loop: Human-centered automated annotation with generative AI. International Journal of Multidisciplinary Futuristic Development, 5(1), 81 to 95. https://doi.org/10.54660/IJMFD.2024.5.1.81-95 Law, R., Leung, R., & Buhalis, D. (2009). Information technology applications in hospitality and tourism: A review of publications from 2005 to 2007. Journal of Travel & Tourism Marketing, 26(5 to 6), 599 to 623. Lawal, O. A., & Oduleye, T. E. (2019). Conceptualizing data driven executive decision systems for strategic financial planning. Iconic Research and Engineering Journals, 3(3). Leys, C., Ley, C., Klein, O., Bernard, P., & Licata, L. (2013). Detecting outliers: Do not use standard deviation around the mean, use absolute deviation around the median. Journal of Experimental Social Psychology, 49(4), 764 to 766. Lipton, Z. C. (2018). The mythos of model interpretability. ACM Queue, 16(3), 31 to 57. Liu, F. T., Ting, K. M., & Zhou, Z.-H. (2008). Isolation forest. In Proceedings of the 8th IEEE International Conference on Data Mining (pp. 413 to 422). Liu, F. T., Ting, K. M., & Zhou, Z.-H. (2012). Isolation-based anomaly detection. ACM Transactions on Knowledge Discovery from Data, 6(1), Article 3. Lopez-Rojas, E. A., & Axelsson, S. (2014). BankSim: A bank payments simulator for fraud detection research. In Proceedings of the 26th European Modeling and Simulation Symposium. Lopez-Rojas, E. A., Elmir, A., & Axelsson, S. (2016). PaySim: A financial mobile money simulator for fraud detection. In Proceedings of the 28th European Modeling and Simulation Symposium (pp. 249 to 255). Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems 30 (pp. 4765 to 4774). Mbonu, I. S., Aliliele, C., Iwuanyanwu, U., & Uzoka, E. (2021a). A conceptual framework for risk based business intelligence architecture in financial technology platforms. International Journal of Multidisciplinary Research and Growth Evaluation, 2(6), 731 to 746. https://doi.org/10.54660/.IJMRGE.2021.2.6.731-746 Mbonu, I. S., Aliliele, C., Iwuanyanwu, U., & Uzoka, E. (2022a). A conceptual framework for AI enabled IT general controls and SOX audit automation processes. Gyanshauryam, International Scientific Refereed Research Journal, 5(5), 384 to 414. https://doi.org/10.32628/GISRRJ2256239 Mbonu, I. S., Iwuanyanwu, U., Aliliele, C., & Uzoka, E. (2021b). A review of VoIP forensic analytics models for financial fraud detection and regulatory compliance monitoring. International Journal of Multidisciplinary Research and Growth Evaluation, 2(6), 711 to 730. https://doi.org/10.54660/.IJMRGE.2021.2.6.711-730 Mbonu, I. S., Iwuanyanwu, U., Aliliele, C., & Uzoka, E. (2022b). A review of data protection impact assessment models in multi cloud financial infrastructure systems. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 8(1), 589 to 623. https://doi.org/10.32628/CSEIT25442 Mbonu, I. S., Iwuanyanwu, U., Uzoka, E., & Oluoha, O. M. (2019). Advances in enterprise log analytics and automated incident response architectures using Python and SIEM platforms. Iconic Research and Engineering Journals, 3(2), 1000 to 1019. https://doi.org/10.64388/IREV3I2-1714915 Medon, J. J., & Oduleye, T. E. (2022). A comprehensive financial reporting model for strengthening compliance and organizational accountability systems. International Journal of Multidisciplinary Research and Growth Evaluation, 3(6), 768 to 777. Medon, J. J., & Oduleye, T. E. (2024). An integrated predictive analytics model for enhancing strategic financial forecasting and decision accuracy. Gyanshauryam, International Scientific Refereed Research Journal, 7(3), 258 to 279. https://doi.org/10.32628/GISRRJ247322 Mendes, T., Cardoso, P. J. S., Monteiro, J., & Raposo, J. (2022). Anomaly detection of consumption in hotel units: A case study comparing isolation forest and variational autoencoder algorithms. Applied Sciences, 13(1), Article 314. https://doi.org/10.3390/app13010314 Miller, T. (2019). Explanation in artificial intelligence: Insights from the social sciences. Artificia