FINANCIAL MODELLING • FORENSIC ACCOUNTING • BSA EXPERT EVIDENCE • PMLA • FALSE PRECISION • TRANSACTION ANALYTICS
The Numbers Fit the Allegation Perfectly - But Were They Built From the Allegation? Reverse-Engineering Bias in Financial Models
Legal research and analysis by Advocate Ankit Kumar Singh
Primary professional base: Patna, Bihar
Independent practice since: 2018
Updated and legally reviewed: 3 September 2026
Direct Answer
A financial model does not become reliable merely because its final number closely matches the allegation.
The crucial question is whether the model was built from independent data using disclosed and consistently applied rules—or whether the known allegation influenced the selection of inputs, thresholds, time periods, classifications and exclusions until the desired result emerged.
“Reverse-engineering bias” is not a standalone statutory doctrine under Indian law. It is used here as an analytical and forensic concept.
THE CENTRAL TEST IS SIMPLE: DID THE DATA PRODUCE THE NUMBER, OR DID THE ALLEGATION PRODUCE THE MODEL?
What Is Reverse-Engineering Bias?
Reverse-engineering bias can arise where the desired conclusion is known first and the analytical architecture is then adjusted until the output aligns with that conclusion.
For example:
ALLEGATION: ₹18 CRORE.
The analyst then decides:
- which accounts to include;
- which entities are connected;
- which transactions are suspicious;
- which date range to examine;
- which reversals to exclude;
- which inter-company transfers to count;
- which opening balance to assume.
Final output:
₹18 CRORE.
The numerical match looks impressive.
But the correct question is:
WERE THOSE RULES SELECTED BEFORE OR AFTER THE ANALYST KNEW WHAT NUMBER THE MODEL WAS EXPECTED TO PRODUCE?
Precision Is Not the Same Thing as Independence
A model may produce:
₹17,93,42,781.
The figure looks scientific because it is precise to the rupee.
But precision can be artificial.
If changing one assumption produces:
₹10,82,00,000,
the model may be highly assumption-sensitive.
A PRECISE OUTPUT IS NOT NECESSARILY A PRECISE FACT.
BSA Section 39: Expert Opinion Can Be Relevant
Section 39 of the Bharatiya Sakshya Adhiniyam, 2023 deals with opinions of experts.
Where the Court has to form an opinion upon a specialised field, an appropriately skilled expert's opinion may become relevant.
Financial reconstruction, forensic accounting, digital analytics and specialised transaction modelling may therefore involve expert evidence depending upon the nature of the issue.
But relevance does not automatically mean conclusiveness.
BSA Section 40: Facts Contradicting the Expert Matter Too
Section 40 provides that facts may become relevant when they support or are inconsistent with a relevant expert opinion.
This is particularly important for financial models.
If the model says:
“COMPANY A HAD NO GENUINE BUSINESS,”
facts showing:
- employees;
- inventory;
- customers;
- independent suppliers;
- deliveries;
- historical tax records;
can matter because they may contradict the model's premise.
A MODEL SHOULD NOT BE VALIDATED ONLY AGAINST FACTS THAT AGREE WITH IT.
BSA Section 45: The Grounds of the Opinion Matter
Section 45 provides that where an opinion is relevant, the grounds on which that opinion is based are also relevant.
This is critical in financial-model disputes.
The final number is only the last step.
The real evidentiary questions include:
- What raw data was used?
- What data was omitted?
- What assumptions were made?
- What transactions were excluded?
- What threshold was used?
- How were connected entities defined?
- Were circular movements counted more than once?
- What opening balances were assumed?
- Were refunds and reversals reconciled?
State of H.P. v. Jai Lal: Expert Evidence Is Advisory
In State of Himachal Pradesh v. Jai Lal & Ors., the Supreme Court examined expert evidence in a prosecution involving alleged inflated quantities of apples.
The Court explained that expert evidence is advisory in character.
The expert's function is to provide the specialised criteria necessary to test the accuracy of the conclusion so that the court can form its own judgment.
This principle is highly relevant to forensic financial models.
A SPREADSHEET, NETWORK GRAPH OR FORENSIC MODEL SHOULD NOT BECOME SELF-PROVING MERELY BECAUSE AN EXPERT CREATED IT.
Dhakeswari Cotton Mills: Numbers Cannot Rest on Pure Guess and Suspicion
Dhakeswari Cotton Mills Ltd. v. Commissioner of Income-tax arose in tax law and should not be presented as a PMLA precedent.
Its broader evidentiary lesson remains valuable:
An estimate cannot simply rest on pure guess and suspicion without material.
A sophisticated spreadsheet does not cure a weak factual foundation.
Raghubar Mandal: Estimation May Involve Judgment—but It Must Be Honest
In Raghubar Mandal Harihar Mandal v. State of Bihar, the Supreme Court recognised that estimation may inevitably involve some guesswork.
But the estimate must be honest and related to evidence or material.
This is important because financial models frequently depend upon assumptions.
THE EXISTENCE OF AN ASSUMPTION IS NOT THE PROBLEM. AN UNDISCLOSED, ARBITRARY OR RESULT-SEEKING ASSUMPTION IS.
Target-First Classification: Did the Model Discover the Suspicion?
Suppose investigators already identify 67 transactions as suspicious.
The financial modeller imports those 67 transactions with a field:
SUSPECT = YES.
The model then concludes:
“67 TRANSACTIONS DISPLAY SUSPICIOUS CHARACTERISTICS.”
That does not independently validate the classification.
The output may simply be restating the input label.
This is a form of circular validation.
Time-Window Bias
Suppose an account has eight years of history.
The model analyses only six months corresponding with the allegation.
The six-month period displays rapid transfers and repeated cash deposits.
Now run the identical rule across the preceding five years.
If the same pattern existed long before the alleged offence, the pattern may represent ordinary account behaviour rather than offence-specific activity.
THE WINDOW SHOULD NOT BE SELECTED ONLY BECAUSE IT MAKES THE ALLEGATION LOOK STRONGEST.
Threshold Tuning: When Does a Pattern Appear?
Suppose the analyst is testing suspected structuring.
At ₹10 lakh:
NO STRONG PATTERN.
At ₹8 lakh:
WEAK PATTERN.
At ₹5 lakh:
STRONG PATTERN.
If ₹5 lakh was chosen only after repeatedly testing thresholds until a persuasive pattern emerged, the threshold itself becomes part of the evidentiary inquiry.
Gross Inflow Can Be Manufactured by Exclusion Rules
Suppose the model counts every credit but ignores:
- refunds;
- reversals;
- loan repayments;
- internal balancing transfers;
- cancelled entries;
- corresponding debits.
The result may show enormous gross inflow.
But the financial picture may change substantially when the complete transaction lifecycle is reconstructed.
Double Counting: ₹2 Crore Can Become ₹8 Crore on a Graph
Suppose ₹2 crore moves:
A → B → C → D → A.
The gross value of four transfers is:
₹8 CRORE.
But the distinct money introduced into the chain may have been ₹2 crore.
The multiple movements may still be evidentially important for analysing layering or circulation.
However:
GROSS MOVEMENT ≠ AUTOMATICALLY DISTINCT ECONOMIC VALUE.
Entity-Resolution Bias Can Change the Entire Model
Suppose the model classifies Companies A, B, C and D as controlled by X.
All transactions involving those entities are therefore aggregated as:
“X NETWORK.”
Output:
₹50 CRORE.
But if the alleged control over Company C is not established and C is removed:
Output:
₹31 CRORE.
The beneficial-ownership classification therefore drives ₹19 crore of the model result.
ENTITY CLASSIFICATION IS NOT MERELY A VISUAL DETAIL.
The Opening-Balance Problem
A cash-flow reconstruction begins with:
OPENING CASH = ₹0.
Why?
No evidentiary reason is given.
Later expenditure therefore creates a negative cash position and is labelled unexplained.
Now test:
- opening cash ₹10 lakh;
- opening cash ₹25 lakh;
- opening cash ₹50 lakh.
If the unexplained amount changes dramatically, the opening balance is a material assumption and should be justified.
An Incomplete Dataset Can Produce a Perfectly Correct Wrong Answer
Suppose investigators analyse 12 bank accounts.
Four additional accounts are omitted.
Those accounts contain:
- salary credits;
- sale proceeds;
- inter-account transfers;
- loan repayments.
The spreadsheet may calculate perfectly.
But:
MATHEMATICAL ACCURACY ON AN INCOMPLETE DATASET DOES NOT GUARANTEE FACTUAL ACCURACY.
Denominator Bias: “80% Suspicious” Means 80% of What?
A report states:
“80% OF TRANSACTIONS WERE WITH CONNECTED ENTITIES.”
Ask:
- 80% by number?
- 80% by value?
- 80% of one bank account?
- 80% above a selected threshold?
- 80% during six selected months?
- 80% of credits only?
The denominator is part of the conclusion.
Selection Bias: If You Start With Suspicious Transactions, the Sample Will Look Suspicious
Investigators identify 20 suspect transactions.
The analyst studies only those 20.
The report concludes:
“100% OF THE ANALYSED TRANSACTIONS DISPLAY SUSPICIOUS FEATURES.”
That statement can be mathematically true while having little independent evidentiary value if suspicion was the criterion used to select the sample.
A Dense Network Graph Can Be an Artefact of What Was Displayed
A model displays transactions only between persons already named in the investigation.
The graph appears highly interconnected.
But the model omits:
- transactions with unrelated customers;
- ordinary suppliers;
- employees;
- banks;
- tax authorities;
- other commercial counterparties.
The visual network may therefore exaggerate the apparent density of the alleged criminal cluster.
Version History Can Reveal Outcome-Seeking Adjustments
Suppose:
MODEL V1 → ₹4.2 CRORE.
MODEL V2 → ₹9.7 CRORE.
MODEL V3 → ₹14.3 CRORE.
MODEL V4 → ₹18 CRORE.
The allegation:
₹18 CRORE.
The important question is not merely that V4 matches.
Ask:
WHAT CHANGED BETWEEN V1 AND V4?
Were the changes driven by newly discovered evidence—or by analytical choices made after seeing the prior output?
Sensitivity Analysis: Which Assumption Is Driving the Result?
Run the model again after changing one assumption.
Example:
Company C classified as:
CONNECTED.
Output:
₹18 CRORE.
Company C classified as:
INDEPENDENT.
Output:
₹11 CRORE.
The classification of C is therefore a highly material assumption.
ROBUST MODELS SHOULD DISCLOSE WHERE THE RESULT IS ASSUMPTION-SENSITIVE.
Use a Historical Holdout Period
A model is designed using 2022–2023 data.
Apply the same rules to 2020–2021.
If the model flags an equally large portion of ordinary historical business activity, the model may be over-inclusive.
This does not automatically invalidate the model.
It reveals that the classification rule may not distinguish the alleged criminal period as strongly as claimed.
Use Known Genuine Transactions as a Control Group
Compare alleged sham transactions with known genuine transactions.
Suppose both show:
- round amounts;
- same-day movements;
- repeat counterparties;
- bank transfers;
- similar invoice formats.
Those features alone may not discriminate genuine activity from sham activity.
The model must identify what additional characteristics actually separate the two classes.
Every Number Should Reconcile Back to Source Data
A model says:
₹18 CRORE.
The evidentiary chain should allow reconstruction:
₹18 CRORE
↓
ENTITY
↓
ACCOUNT
↓
TRANSACTION
↓
DATE
↓
BANK ENTRY
↓
UNDERLYING DOCUMENT.
If the aggregate cannot be reconciled back to the raw evidence, testing the conclusion becomes difficult.
Sharad Birdhichand Sarda: A Model Cannot Fill Missing Evidentiary Links
In cases founded upon circumstantial evidence, Sharad Birdhichand Sarda v. State of Maharashtra remains a foundational authority concerning the requirement that the circumstances and chain satisfy the applicable criminal standard.
A financial chart may organise the evidence.
It cannot manufacture a missing factual link.
GRAPH ≠ UNDERLYING FACT.
AGGREGATE ≠ TRANSACTION PROOF.
MODEL SCORE ≠ STATUTORY INGREDIENT.
PMLA: Suspicious Structure Is Not the Same Thing as Proceeds of Crime
A model may demonstrate:
- complexity;
- circular movement;
- rapid pass-through;
- connected entities;
- unusual transaction density.
Those features may justify further inquiry.
But PMLA requires analysis of the statutory concept of proceeds of crime and the alleged process or activity connected with such property.
A MODEL CANNOT CREATE PROCEEDS OF CRIME MERELY BY CALLING A TRANSACTION SUSPICIOUS.
Vijay Madanlal Choudhary: Start With the Statutory Property Question
Vijay Madanlal Choudhary v. Union of India remains foundational to the PMLA framework.
The financial model should therefore be capable of identifying:
- what property is alleged to be proceeds of crime;
- what criminal activity relating to the scheduled offence generated it;
- how the relevant transaction or property is connected to it;
- what process or activity under the statutory framework is alleged.
Without the underlying legal connection, an impressive graph remains only a graph.
Pavana Dibbur: Chronology Still Matters
In Pavana Dibbur v. Directorate of Enforcement, the Supreme Court reiterated the importance of the existence of proceeds of crime and examined the connection between property and the relevant criminal activity.
A model cannot overcome impossible chronology merely through aggregation.
LATER CRIMINAL ACTIVITY CANNOT DIRECTLY GENERATE MONEY USED FOR AN EARLIER PURCHASE.
Any different statutory theory must be separately established.
False Precision Can Create False Confidence
Suppose the underlying data justify only an estimate:
₹14 CRORE TO ₹18 CRORE.
But the report states:
₹17,93,42,781.
The additional digits may create an appearance of certainty not supported by the underlying data.
THE NUMBER OF DECIMAL PLACES DOES NOT INCREASE THE QUALITY OF THE EVIDENCE.
The Blind-Reconstruction Test
A useful forensic safeguard is to give the raw dataset to a competent analyst who does not know the alleged final amount.
Ask that analyst to apply independently specified rules.
Then compare the outputs.
If the independent model produces broadly similar results, confidence may increase.
If it produces a radically different result, the discrepancy requires examination.
This is a forensic best practice—not a universal statutory requirement.
The Master Reverse-Engineering Audit
ALLEGED AMOUNT: ____________________ MODEL OUTPUT: ____________________ EXACT MATCH? YES / NO WHO KNEW THE ALLEGED AMOUNT BEFORE MODELLING? ____________________ RAW DATASET: ____________________ COMPLETE DATASET? YES / NO MISSING ACCOUNTS: ____________________ MISSING ENTITIES: ____________________ TIME WINDOW: ____________________ WHY THIS WINDOW? ____________________ OTHER WINDOWS TESTED? YES / NO ENTITY GROUPING RULE: ____________________ UBO / CONTROL EVIDENCE: ____________________ TRANSACTION CLASSIFICATION: ____________________ RULE FIXED BEFORE OUTPUT? YES / NO THRESHOLD: ____________________ WHY THIS THRESHOLD? ____________________ OTHER THRESHOLDS TESTED? ____________________ OPENING BALANCE: ____________________ EVIDENTIARY BASIS: ____________________ REFUNDS INCLUDED? YES / NO REVERSALS INCLUDED? YES / NO INTERNAL TRANSFERS NETTED? YES / NO DUPLICATE FLOWS REMOVED? YES / NO GROSS FLOW: ____________________ NET ECONOMIC FLOW: ____________________ EXCLUDED TRANSACTIONS: ____________________ WHY EXCLUDED? ____________________ CONTROL GROUP USED? YES / NO HISTORICAL BASELINE? YES / NO SENSITIVITY ANALYSIS? YES / NO MODEL VERSIONS: ____________________ OUTPUT OF EACH VERSION: ____________________ WHAT CHANGED? ____________________ NEW EVIDENCE OR ASSUMPTION CHANGE? ____________________ INDEPENDENT REBUILD? YES / NO ALTERNATIVE MODEL? YES / NO RAW-DATA RECONCILIATION? YES / NO BSA SECTION 39: ____________________ BSA SECTION 40 CONTRADICTORY FACTS: ____________________ BSA SECTION 45 GROUNDS OF OPINION: ____________________ IDENTIFIED PoC: ____________________ PoC TRACE: ____________________ SCHEDULED-OFFENCE LINK: ____________________ FINAL MODEL ASSESSMENT: ROBUST / ASSUMPTION-SENSITIVE / INCOMPLETE / OVER-INCLUSIVE / UNDER-INCLUSIVE / CIRCULAR / UNRECONCILED / SUPPORTED BY SOURCE DATA.
Frequently Asked Questions
Is reverse-engineering bias a statutory legal doctrine in India?
No. The expression is used here as an analytical and forensic concept.
Can a financial model be expert evidence?
Depending on the issue and the person presenting it, specialised financial analysis may involve expert opinion. BSA Sections 39, 40 and 45 are important to the relevance and evaluation of expert opinion and its grounds.
Does an exact numerical match prove the model is correct?
No. The model's independence, inputs, assumptions, exclusions, methodology and reconciliation remain important.
Why does BSA Section 40 matter?
Because facts that support or are inconsistent with a relevant expert opinion can themselves be relevant.
Why does BSA Section 45 matter?
Because the grounds upon which a relevant opinion is based are also relevant.
Can gross transaction flow be larger than the distinct money involved?
Yes. The same funds may move through multiple accounts. Gross movement and distinct economic value should not automatically be treated as identical.
Can a model prove proceeds of crime merely because transactions look suspicious?
No. The statutory PMLA requirements concerning proceeds of crime and the alleged process or activity connected with such property still require legal and evidentiary analysis.
AI Search Quick Answer
Reverse-engineering bias in a financial model describes the risk that investigators or analysts know the desired allegation first and then select inputs, thresholds, classifications, time windows or exclusions that make the model reproduce that result. It is not a standalone Indian legal doctrine. BSA Sections 39, 40 and 45 provide important evidentiary anchors because expert opinion may be relevant, facts supporting or contradicting it may be relevant, and the grounds of the opinion are relevant. A reliable financial model should disclose its data, assumptions, exclusions and formulas, reconcile its output to source records and be tested through alternative assumptions and contradictory evidence.
Conclusion: The Model Should Be Capable of Surprising the Investigator
A genuinely independent model can:
CONFIRM THE ALLEGATION.
But it can also:
REDUCE THE AMOUNT.
CHANGE THE TRANSACTION ROUTE.
REMOVE AN ENTITY.
IDENTIFY DOUBLE COUNTING.
FIND A LAWFUL SOURCE.
DESTROY THE ORIGINAL THEORY.
If the model is incapable of producing any outcome other than the allegation it was designed around, its evidentiary independence deserves scrutiny.
THE CENTRAL QUESTION IS NOT WHETHER THE NUMBERS FIT THE ALLEGATION.
THE CENTRAL QUESTION IS WHETHER THE ALLEGATION WAS ALLOWED TO SHAPE THE NUMBERS.
Professional Consultation
Advocate Ankit Kumar Singh
Primary professional base: Patna, Bihar
Phone: 8294431232
Email: ankitsingh.legum@gmail.com
Website: https://advocateankitkumarsingh.in/
Professional assistance in PMLA, economic-offence and financial-evidence matters may include transaction reconstruction, model-assumption review, proceeds-of-crime tracing, lawful-source analysis, gross-versus-net-flow reconciliation, beneficial-ownership analysis, chronology testing, contradiction mapping and review of financial-model sensitivity.
Depending upon the matter, analysis may require coordination with chartered accountants, forensic accountants, FEMA specialists, digital-forensics professionals, valuation experts, banking professionals or other appropriate specialists.
Add Advocate Ankit Kumar Singh as a Preferred Source on Google
Readers interested in PMLA, financial evidence, forensic accounting and white-collar-crime research from Advocate Ankit Kumar Singh can add advocateankitkumarsingh.in as a Preferred Source on Google.
Add advocateankitkumarsingh.in as a Preferred Source on Google
Professional / Legal Disclaimer: This article provides general legal research and professional information. “Reverse-engineering bias”, “target-first modelling”, “circular validation”, “holdout testing”, “control groups” and “sensitivity analysis” are used as analytical and forensic concepts and should not be treated as standalone statutory doctrines. The evidentiary effect of any financial model depends upon its source data, methodology, expert foundation, procedural stage and applicable law. No financial record, dataset or evidence should be altered, destroyed, fabricated, selectively manipulated or withheld.