A.5.5 Failure Modes & Practice

Basic Prerequisites: 02 · The Accrual Anomaly through 04 · Red Flags & Shenanigans.

---

1. Intuition & Practical Objective

Pages 02–04 gave you instruments: the accrual ratio, the discretionary-accrual models, the archetype checklist. This page is the honesty layer - the ways each instrument fails, each failure traced to a first principle rather than presented as a gotcha. If the hub’s checklist is the engine, this page is the quality control on the engine itself.

The deepest principle, stated once and used everywhere below:

Every quality measure is a difference between two accounting numbers, and accounting numbers are jointly produced by the same firm under the same rules. So the measures share blind spots: manipulation that moves both sides of the difference equally is invisible, and manipulation that reverses is self-erasing.

Three consequences follow directly and are the spine of this page: (1) accruals reverse, so a single-year measure is timing-luck; (2) everything booked to cash cancels out of balance-sheet accruals; (3) consolidation and comparability changes make time-series screens read structure as quality.

---

2. Mathematical Ground Truth & Derivations

Reversal - why the accrual identity is a difference, not a level. Accruals sum to zero over the life of the firm (you can defer recognition, never eliminate it). So if firm $F$ injects $X$ of accruals in year 1, year 2 carries $-X$:

$${\text{NI}}_1={\text{CF}}_1+X,{\text{NI}}_2={\text{CF}}_2-X,{\text{NI}}_1+{\text{NI}}_2={\text{CF}}_1+{\text{CF}}_2.$$

The cumulative earnings are unchanged; only the timing moved. Therefore the only accrual measure that is robust to reversal timing is the cumulative one:

$$\text{cumulative CFO/NI}=\frac{{\sum }_{t=1}^T{\text{CFO}}_t}{{\sum }_{t=1}^T{\text{NI}}_t}<1\Rightarrow \text{earnings were booked that never became cash.}$$

A single-year accrual ratio cannot distinguish “manipulating upward this year” from “reversing last year’s manipulation” - they have opposite signs and the same cause. This is why the trend and the cumulative version outrank the level.

Blindness to cash-side games. Write the Sloan measure as a sum of differences of balance-sheet changes. Fabricate a cash sale of amount $X$: both $\Delta CA$ (via cash) and $\Delta Cash$ rise by $X$, so

$$(\Delta CA-\Delta Cash)\text{is unchanged by }X.$$

The accrual measure is algebraically incapable of seeing it. The same cancellation kills the CFO/NI ratio as a detector once the fake cash is booked as operating inflow. Consequence: the balance-sheet toolbox detects timing manipulation; it does not detect fabrication.

Comparability and the ratio’s denominator. A screen compares this year’s ratio to last year’s - which assumes the two years are comparable. An acquisition breaks that assumption:

$$\text{CFO}\uparrow \text{ by the acquired firm’s cash flows},\text{NI}\uparrow \text{ by the acquired firm’s earnings},\frac{\text{CFO}}{\text{NI}}\text{ reads as "improving quality"}$$

even when the acquirer’s own operations are deteriorating. The same logic applies to $\Delta CA$, goodwill, and any balance-sheet stock after a material transaction. The practice rule is mechanical: flag the year a material acquisition or divestiture occurred, and never read a quality trend straight through it.

Composite scores reward improvement, not level. A binary signal like $\Delta \text{ROA}>0$ is scale-free - it cannot tell whether the firm went from $+15\%$ to $+16\%$ or from $-15\%$ to $-2\%$. Both score 1. Any composite built from directional signals inherits this and will rank a deeply distressed firm that is merely deteriorating less slowly alongside a genuinely strong one.

---

3. Computational Implementation - four ways the screens lie

Stdlib only. Four failure modes, each demonstrated with the minimum arithmetic that produces the failure, so the mechanism is undeniable rather than asserted.

Mode 3 is the one to remember: the $+50$ of fabricated cash revenue moves the Sloan accrual expression by exactly zero. The instrument is not weak here - it is structurally blind.

---

4. Failure Modes & First-Principles Breakdowns

The catalogue, each tied to its first principle.

1. Timing luck mistaken for quality (first principle: accruals reverse). A one-year accrual ratio conflates manipulation with the reversal of last year’s manipulation. Rule: use trend and cumulative measures; never act on a single year’s accrual ratio. → 02 · The Accrual Anomaly.
2. Structural blindness to cash-side fabrication (first principle: the difference cancels). Fabricated cash revenue leaves balance-sheet accruals untouched and improves CFO/NI. Rule: for accounts with cash-side exposure, use external evidence (channel checks, customer filings, returns, receivables aging) - the ratio will never tell you.
3. Comparability breaks (first principle: time-series ratios assume a stable entity). Acquisitions, divestitures, FX, and accounting-standard changes (IFRS/ASC adoptions, lease capitalisation) all reset the level. Rule: mark the break year; rebuild the series on a constant-perimeter basis or skip the comparison. → 04 · Red Flags & Shenanigans.
4. The proxy you chose determines the answer (first principle: “quality” is not one construct). Accruals-based, persistence-based, smoothness-based and restatement-based measures rank firms differently - and can rank them oppositely. Rule: name the construct before you screen (Dechow, Ge & Schrand’s whole point) and never mix proxies into a single “quality score” without saying what it means.
5. Extreme-performance contamination (first principle: benchmarks under-control for performance). Poorly performing firms have extreme accruals for legitimate reasons, and every discretionary-accrual model over-rejects in those samples. Rule: add a performance control, or restrict the sample, or report that you did not. → 03 · Detecting Earnings Management.
6. Look-ahead and survivorship (first principle: data hygiene). Using restated fundamentals to test a quality screen bakes the outcome into the input; testing on today’s live firms drops every delisted fraud. Rule: point-in-time, as-reported data, delisting returns included. → Data Sources & Corporate Data.
7. Directional composites reward non-death (first principle: binary signals are scale-free). “Improving” from a catastrophic base scores the same as “improving” from excellence. Rule: pair every binary signal with a level condition (e.g. ROA > 0, not merely rising).
8. Discretion is not fraud (first principle: a residual is a residual). Discretionary accruals measure deviation from your model - which includes legitimate judgement, differing policy, and superior growth. Rule: treat every quality flag as a question, and state the innocent explanation you ruled out.

---

5. References

• Dechow, Patricia, Ge, Weili & Schrand, Catherine: “Understanding Earnings Quality: A Review of the Proxies, Their Determinants and Their Consequences” (JAE, 50(2–3), 344–401, 2010)
• Sloan, Richard G.: “Do Stock Prices Fully Reflect Information in Accruals and Cash Flows About Future Earnings?” (TAR, 71(3), 289–315, 1996)
• Dechow, Sloan & Sweeney: “Detecting Earnings Management” (TAR, 70(2), 193–225, 1995)
• Schilit, Perler & Engelhart: Financial Shenanigans (McGraw-Hill, 4th ed., 2020)
• Chan, Jegadeesh & Lakonishok: “Earnings Quality and Stock Returns” (JF, 61(2), 769–806, 2006)
• Mulford, Charles W. & Comiskey, Eugene E.: The Financial Numbers Game (Wiley, 2002)

---

6. Connected Graph Bridges

• Back: 03 · Detecting Earnings Management · 04 · Red Flags & Shenanigans · Index Hub
• Forward: 06 · Advanced Extensions (the Beneish M-score, and proxies that avoid some of these traps)
• Parallel defense layer: Ratio Red Flags
• Data hygiene and incentives: Data Sources & Corporate Data · Capital Structure & Corporate Finance (debt-covenant and bonus-plan incentives that trigger management)