The retrospective method

Revised practical article; source details and figures are dated to their cited filing or event, with current evidence reviewed through 24 September 2026.

After a company fails, its final crisis can make the ending seem obvious. That feeling is dangerous. The reader now knows which supplier went unpaid, which loan fell due, and which rescue did not arrive. A lender or owner three years earlier did not know those outcomes with certainty.

Fischhoff’s peer-reviewed experiments found that knowledge of an outcome changes how people judge its prior likelihood and how they remember what they knew before [1]. A failure retrospective should therefore keep two timelines: the public information that existed at each decision date, and the later information revealed in administration, court proceedings or liquidation.

Start with the decision date

For every important claim, record:

A leaked presentation may be contemporaneous but unaudited. A court judgment is authoritative about what the court decided and can summarize a party’s submissions, but a party’s claim in a judgment is not automatically a proven fact. An administrator report may give the best available later view of creditors or assets, but it is not what an investor necessarily knew at the earlier date. A media article can help locate a primary document without replacing it.

Rebuild the event sequence before assigning causes

Construct a dated chronology that can include:

• changes to the operating model, footprint, customer terms, or supplier terms;
• audited financial statements and late filings;
• new borrowing, extensions, missed payments, or refinancing;
• changes in owner or management control;
• publicly announced funding, asset sales, layoffs or restructurings;
• creditor petitions, administration appointments, court rulings, liquidation orders, and later distributions.

Then connect each date to operating signals: cash runway, receivables ageing, inventory turns, gross contribution, customer concentration, fixed commitments, capex, debt maturities, and working-capital needs. Use only the data available at that time for the early-warning analysis. Place later events in a separate column so they do not leak backward into the forecast.

A warning system is a conversation, not a score

Peer-reviewed distress research shows why a useful model needs time-aware data and more than one accounting ratio. Shumway argues that a hazard model handles the timing of bankruptcy more appropriately than a single-period classification and finds market variables can add information [2]. Campbell, Hilscher and Szilagyi study distress using accounting and market indicators and distinguish failure risk from subsequent stock returns [3]. These studies use public-company datasets and defined failure labels. They do not supply a calibrated warning score for Kenyan family businesses or startups. Merton’s contingent-claims approach offers a separate way to think about default risk when the market value and volatility of a firm’s assets and its debt structure can be estimated [4]. Those inputs are rarely available for private Kenyan businesses, so the framework should not be dressed up as a local probability model.

For an MSME, use a small set of observable indicators to start questions, not to issue an automatic “safe/unsafe” verdict. Examples:

• the cash forecast’s next low point moves closer or below the reserve floor;
• the largest customer’s receipt is delayed or disputed;
• a supplier shortens payment terms or requires cash in advance;
• inventory grows faster than sales or becomes hard to sell;
• gross contribution weakens while fixed costs stay in place;
• scheduled debt repayments exceed the forecast’s available cash;
• the owner’s key duties, bank access or supplier relationships have no backup.

For each signal, name the decision it changes. If a customer owes a material amount, follow up and update the receipt date. If stock is growing, test demand and ageing by product. If a loan matures in eight months, begin a refinancing or payoff discussion before the last month. The signal must be specific enough to act on and auditable enough to learn from.

Test against false alarms and missed cases

Every warning system produces false alarms and misses. If we only study companies that failed, we will not know which warning signs were common among businesses that recovered. Define the event first: late payment, breach, administration, court liquidation order, insolvency filing, business closure, or owner exit? These are not one outcome. Select a forecast horizon, preserve cases that survived, and disclose the limits of the sample.

A legal process is also not an operating diagnosis. Administration can be used to pursue a rescue or better outcome for creditors; it does not mean liquidation has already occurred. A company under administration may continue operating, restructure or sell assets. A liquidation order is a different legal milestone. State the exact event and source.

The practical retrospective test

For every major alleged cause, ask:

1. What is the strongest primary record for it?
2. Was the record available before the event or only after it?
3. Is it describing cash pressure, accounting performance, contractual default, or a legal outcome?
4. Are there plausible alternative explanations?
5. Would the proposed early warning have been visible to an owner with ordinary records at that time?
6. Did it produce a decision soon enough to matter?

If a claim cannot pass this test, label it as a hypothesis or unresolved question. A good retrospective does not need a villain or a single ratio. It needs to show what was observable, what action could reasonably have followed, and where the evidence runs out.

From an indicator to an owner decision

A useful warning sheet has four parts: the measurement, its trigger, the owner of the response, and the time available to act. “Receivables are rising” is only an observation. A more useful statement is: “The three largest overdue invoices are now KSh 1.8 million, equal to six weeks of payroll; the credit lead will call the buyers today, pause new exposure above the approved limit, and refresh the cash forecast on Friday.” The amounts and thresholds must come from the business’s own records and obligations. The example illustrates the form of a decision, not a universal safe limit.

Separate leading indicators from outcomes. A supplier asking for cash on delivery, a lost credit line, or a new covenant waiver can precede a missed payment. A bounced payment, court petition, or administration appointment is usually a later outcome. A retrospective should ask whether the earlier signal was visible, whether it was recorded, and whether anyone had authority and time to respond. If no one was assigned the decision, a technically accurate dashboard may still have had no practical value.

The forecast itself needs a short, honest horizon. A weekly cash schedule can list opening unrestricted cash, confirmed customer receipts, committed purchases, wages, rent, taxes, debt service and a minimum operating reserve. Put uncertain receipts in a separate column with a probability or evidence note; do not quietly treat a sales promise as cash. Run a base case and one or two adverse cases, such as a key customer paying two weeks late or a supplier withdrawing terms. The point is not to predict the exact failure date. It is to find which assumption would force a decision first.

After each review, preserve the forecast version and the actual result. Compare the amount and date of expected receipts and payments with what happened. This reveals whether repeated errors came from late customer data, optimistic sales assumptions, unrecorded obligations, or a genuinely unpredictable event. A warning system should improve its inputs and ownership as well as its thresholds.

What a retrospective can fairly conclude

A sound account distinguishes a mechanism from proof of causation. If a retailer has a large overdue supplier balance, that can indicate a cash squeeze and can threaten replenishment. To claim that the arrears caused closure, however, we would need evidence about timing, the supplier’s response, other financing, store economics and plausible alternatives. When documents show sequence but not the counterfactual, write “consistent with” or “may have contributed,” then say what evidence would decide the issue.

This restraint does not make the analysis timid. It helps owners act on claims they can verify: cash dates, inventory ageing, customer concentration, contract renewals, debt maturities and who controls payment access. It also makes cases comparable. A company with an administration order, one with a missed loan payment and one that closed without a formal filing should not be counted as identical failures merely because a news article used the word “collapsed.”

Build the case file so someone else can reproduce the timeline

A retrospective becomes practical when it can be rebuilt from records rather than memory. Start a case register with one row per legal entity, trading name, facility or material business unit. Record registration number if available, ownership, trading period, jurisdiction, administrator or liquidator, court case number, and whether the entity continued trading. Add a relationship map that distinguishes parent, subsidiary, shareholder, lender, supplier, brand and asset owner. A shared name or director does not prove that two companies had the same cash or liability.

Then create an evidence ledger. Each item should include source title, publisher or custodian, source date, date it became public, period covered, entity, page or paragraph, evidence type, exact claim supported, limitations, and a link or file path. Preserve a short excerpt or factual note rather than copying pages of text. A judgment may establish that the court made an order on a date and may summarize a report. It does not automatically independently verify every allegation quoted from a party. A press release can show what management claimed at the time. A later administrator report may show what records were available after control changed. Those sources answer different questions.

Use four confidence labels:

• Established by primary record: the document directly states the fact, such as a court order date or an audited balance.
• Attributed statement: a named person or report made the claim, but the underlying evidence is not independently available.
• Analyst reconstruction: a value is calculated from stated inputs using a visible method.
• Open question: the available record does not support a conclusion.

These labels prevent a neat narrative from flattening different evidence into one voice. They also help later researchers identify the missing page or file that would upgrade a tentative claim.

Keep a reporting vector, not a headline number

A row containing “KES5.2bn, 2021” is not enough to plot a trend or compare it with a later balance. The amount needs a reporting vector: the legal entity; the measure; amount and unit; period or date the measure covers; date of the underlying event; date the record was prepared; date the record became public; source and page; and status as a filed fact, attributed claim, calculation or estimate. Add the retrieval date so a later researcher can identify which version was checked. Store “not stated” in a missing field. Do not borrow a date from another column to make a blank look complete.

The Nakumatt court record shows why. A ruling delivered on 11 March 2021 summarized an administrator update prepared on 1 February 2021. The administrator was reported to have raised KSh5,218,737,409; more than KSh3.5bn had been paid to creditors, including KSh766,641,309 to landlords [8]. Those are three distinct measurements. The landlord payments sit inside the creditor-payment description, so adding them would double count. The court summary does not provide a transaction-level schedule, the full period covered by “raised,” or individual receipt and payment dates. The report date cannot be treated as the date that all the money was raised or paid.

In the ledger, write the first amount as an administrator-reported cumulative figure with its coverage period marked unknown. Record the second as a lower bound, not an exact KSh3.5bn. Record the landlord amount as an included component, not a fourth cash flow. Give each a source locator and a separate date the court made the summary public. A chart should either show these as dated report observations—with a label that the underlying period is unknown—or keep them out of a transaction-timing series. It should never place the full amount on 1 February merely because that is the report date.

This simple discipline is useful for investor rounds, debt facilities, court claims and operating results alike. A financing announcement, signed agreement, close, drawdown and bank receipt are separate states. A judgment can repeat a petitioner's claim without converting it into an admitted liability. A reported annual figure has a financial period that may end months before publication. The Evidence Vectors sheet in the workbook is designed to keep those dates and evidence states visible; the article uses its rows only for the claim each source can actually support.

Rebuild cash pressure month by month where the record permits

Annual accounts are often too coarse to show the order in which a business ran out of options. If monthly management accounts or bank statements are available, calculate unrestricted cash by week or month, collections, payroll, critical supplier payments, statutory obligations, debt service and owner injections. Reconcile opening to closing bank cash and identify restricted, pledged or client money. If only annual figures exist, do not invent monthly runway; show the broad annual trend and identify the missing data.

A private-company runway should be based on an operating cash forecast, not a headline cash balance divided by recent monthly burn. A cash balance may include customer deposits, restricted funds, borrowed money or proceeds from an asset sale. Burn can change with seasonality, launch spending, a large stock purchase, collections and creditor terms. Reconstruct the period's movement first. Then calculate how much cash remained after obligations already due and after committed spending.

For each period, separate operating receipts from financing. New equity, debt, grants, owner advances, supplier credit and asset-sale proceeds are different sources with different claims. A new round can extend runway without improving gross margin or contribution. A facility can delay default while increasing future repayment pressure. A supplier extension can preserve cash now while leaving an overdue payable. The article should say what the financing achieved and what risk it moved forward.

When only balance-sheet snapshots are available, estimate a broad range and state the assumptions. The lower bound might include confirmed cash and known immediate obligations; the upper bound might include uncollected receivables only to the degree supported by subsequent receipts and exclude disputed claims. Do not label the result “runway” without a time horizon and a burn rate consistent with the cash scope.

Separate early warnings from late outcomes

An owner can use a warning system only if the indicator arrives early enough for an action. Classify each indicator as leading, coincident or lagging. A weekly fall in minimum forecast cash may be leading. A supplier reducing terms can be an early signal or a routine renegotiation. A returned payment is a serious event but may already be late. A court petition or administration appointment usually belongs in the event chronology, not the early-warning dashboard.

For each signal, test the operational chain:

1. What changed in the underlying evidence?
2. Who received the information and on what date?
3. Who had authority to act?
4. What realistic action was available then?
5. How long would that action take to affect cash?
6. Did the business have enough time and capacity to carry it out?

This chain matters because dashboards do not rescue companies by themselves. A sales manager may know that a major account is disputing invoices but lack authority to stop new credit. A founder may know a loan is coming due but wait because the lender gave a verbal assurance. A board may receive a forecast after an irreversible order has been placed. The retrospective should test the organization around the indicator, not only the indicator.

A practical example: a cash warning that was actionable

Imagine a fictional food distributor, Coastal Supply Ltd. At the end of March, it has KSh 6.0 million of unrestricted bank cash, KSh 12.0 million of invoices due from customers over the next month, payroll of KSh 2.4 million, supplier commitments of KSh 5.0 million, tax of KSh 0.8 million, and debt service of KSh 1.2 million. The 13-week forecast shows a minimum balance of KSh 0.3 million in week five. One customer owes KSh 4.0 million, has disputed delivery of two orders and historically pays 30 to 45 days late.

The warning is not simply that cash is “low.” It is that the forecast relies on a specific disputed KSh 4.0 million receipt before supplier and payroll payments fall due. The owner can ask the customer to confirm accepted delivery, send proof of delivery, collect the undisputed portion, negotiate a partial payment plan, pause further credit above an approved exposure limit, and discuss a staged supplier payment before the due date. The company can also defer a discretionary equipment purchase if doing so does not create a safety or reliability risk.

At the next review, compare each action with cash outcome and customer contribution. If the customer pays the undisputed amount but disputes the rest, update future credit terms and the cash forecast. If the supplier accepts a documented schedule without penalty or lost supply, record that. If the company freezes all customer orders and loses a profitable reliable account, the warning response has protected cash at too high a cost. A practical system evaluates decisions by their effects on cash, gross contribution, service, legal obligations and relationships.

The fictional example illustrates response design only; it is not a factual account of a Kenyan company or a universal threshold. A real case should use the source ledger to show when the owner knew the dispute, whether the cash forecast existed, and what alternatives were contractually and operationally available.

Test whether the warning would have survived a false-alarm review

After a failure, an analyst can easily identify many signals that looked ominous in hindsight. The challenge is to ask whether the same measure would have produced too many false alarms among businesses that survived. Use both failed and continuing companies where the data allow. Preserve the cases that recovered, refinanced or changed strategy. Record firms that faced late customers, supplier pressure or high leverage but did not fail. This lets the researcher distinguish a risk factor from a deterministic cause.

Define the unit of analysis and outcome before counting cases. A company can miss a payment but remain in business; an entity can enter administration while stores continue operating; a brand can disappear while the legal company is still under a process; a founder can sell the business without insolvency. Do not combine these into one “failure” label without explaining the choice.

If the sample is small, avoid fitting a complex statistical model. Use a case table and document the observable pattern. If the researcher does estimate a threshold, report the number of cases, time period, data missingness, look-ahead bias, false positive rate and missed-case rate. Shumway's hazard-model approach is relevant because distress risk changes over time and covariates can vary [2]. Campbell, Hilscher and Szilagyi use public-market and accounting signals in a different sample and separate distress from returns [3]. These are methodological reference points, not plug-in coefficients for local private firms.

From retrospective findings to an operating control

The final product should translate each credible warning into a control that a living business can use. A control needs a measure, source, frequency, threshold, response owner, action, deadline and escalation path. For instance, “largest overdue receivable exceeds the amount the 13-week forecast can absorb” can trigger a same-day collection review, credit approval pause for additional exposure, and customer escalation. The threshold derives from actual liquidity and contractual terms rather than a generic industry ratio.

Match the action to the warning type. A collection problem may require document repair or customer contact. A margin problem may require SKU-level contribution analysis and repricing. Inventory ageing may call for a clearance or purchasing hold, but not if the inventory is a critical spare or a confirmed seasonal product. Refinancing risk needs a lender pack months before maturity. Governance and key-person risk may require dual bank access, succession documentation or board escalation.

Record the decision and review date. If an exception is overridden, write who approved it, why, what evidence supports it, when it expires and what would reverse the decision. A control that can be overridden silently is not a reliable warning system.

A fair retrospective has a useful ending

Close each case with four statements:

• What the public record establishes: dates, orders, filed balances, named appointments or disclosed payments.
• What the record suggests but does not prove: mechanisms consistent with stress, such as rising overdue creditors or cash-consuming growth.
• What remains unknown: monthly cash, contract terms, inventory realizability, owner withdrawals, related-party exposure or lender negotiations.
• What a similar owner can do now: forecast low cash, age receivables and stock, schedule maturities, document decision rights and test alternate actions.

The value is in giving an owner a practical early action while preserving the limits of the evidence. This approach avoids both hindsight certainty and vague caution: it says what could have been measured, what decision might have followed, and what outcome should be tracked to learn whether the intervention worked.

Preserve the analyst's own decision record

A complete retrospective also records what the analyst recommended, what evidence was missing, how confident the estimate was, and what later event would have changed the view. This allows a team to distinguish a weak operating signal from a reasonable judgment made with incomplete information. The purpose is to improve the next cash forecast and response, not to reward hindsight for predicting an outcome that was unknowable at the time.

A retrospective should also state the decision boundary. A credible warning may justify calling a customer, pausing a new commitment or seeking advice without proving that insolvency was inevitable. The article should distinguish an operating intervention from a legal conclusion, then note when the evidence became sufficient to escalate.

The strongest output is a dated record that says what was knowable, when the response could still help, and what later evidence changed the interpretation. That record lets an owner act promptly without turning a warning into a verdict.

The review closes with a specific update condition: name the filing, cash report, court document or operating measure that would change the current view, and schedule when to check it. That keeps the retrospective useful after publication.

The outcome note should also say who will check the next record and by what date. A signal that no one owns can be obvious in hindsight and useless in the moment.

A future update should preserve the earlier data cut while adding new administrator reports, judgments or recoveries as later evidence. This makes the chronology auditable instead of rewriting the past with information learned afterward.

Workbook map

The Project 3 workbook includes a legal-entity register, decision-date timeline, evidence ledger, funding-raised register, fictional 13-week cash and CCC case, warnings and actions, and Sources and Checks. Evidence Vectors separates underlying event date, source-record date and public judgment/report date, while identifying whether an entry is a legal fact, administrator/creditor claim, secondary report, calculation, scenario or gap. It separates court-recorded claims from analyst sums and secondary reporting. The cash model is a teaching example, not historical company cash flow, and the warning sheet asks for a company-specific trigger, owner, action and response time rather than a universal failure score.

Technical appendices

The worked calculations and source boundaries below extend this article's examples. They keep assumptions and unresolved inputs visible so readers can reproduce the steps without mistaking an illustration for an issuer or sector estimate.

• Appendix 01 - A reproducible retrospective and warning test
• Appendix 02 - Court-record evidence, entity boundaries and open claims

References

[1] B. Fischhoff, “Hindsight Is Not Equal to Foresight: The Effect of Outcome Knowledge on Judgment Under Uncertainty,” Journal of Experimental Psychology: Human Perception and Performance, vol. 1, no. 3, pp. 288–299, 1975, https://doi.org/10.1037/0096-1523.1.3.288.

[2] T. Shumway, “Forecasting Bankruptcy More Accurately: A Simple Hazard Model,” The Journal of Business, vol. 74, no. 1, pp. 101–124, 2001, https://doi.org/10.1086/209665.

[3] J. Y. Campbell, J. Hilscher, and J. Szilagyi, “In Search of Distress Risk,” The Journal of Finance, vol. 63, no. 6, pp. 2899–2939, 2008, https://doi.org/10.1111/j.1540-6261.2008.01416.x.

[4] R. A. C. Merton, “On the Pricing of Corporate Debt: The Risk Structure of Interest Rates,” The Journal of Finance, vol. 29, no. 2, pp. 449–470, 1974, https://doi.org/10.1111/j.1540-6261.1974.tb03058.x.

[7] Kenindia Assurance Co Ltd v Administrator, Nakumatt Holdings Ltd (UA) & another, [2025] KEHC 12545 (KLR), ruling delivered 9 September 2025, https://new.kenyalaw.org/akn/ke/judgment/kehc/2025/12545/eng@2025-09-09.

[8] In re Nakumatt Holdings Limited (Under Administration), [2021] KEHC 8603 (KLR), ruling delivered 11 March 2021; accessible reproduction of the court ruling, https://sheriahub.com/cases/ke/caselaw/in-re-nakumatt-holdings-limited-under-administration-2021-eklr.pdf.

Companion resources

The workbook and technical appendices are available with research membership. The video production guide maps this article to a planned episode; a published video is not yet available.