Editorial Review for Finance Writing
A re-read is not a review. A review is six passes, each looking for one class of defect. Reading for everything at once means finding nothing, because the interesting failures hide behind fluent prose.
Run the passes in order. Do not line-edit before the evidence audit — you will polish sentences you are about to delete.
Pass 1 — Evidence audit
For every factual claim in the draft, ask: where is this from?
| Defect | Symptom | Fix |
|---|---|---|
| Unsourced number | A percentage with no attribution | Find the source or cut the number |
| Anonymous authority | "studies show", "experts say", "research finds" | Name the study, or cut |
| Missing trial count | A backtest Sharpe with no N | Add N, or state that it is unreported and treat that as the finding |
| Unnamed benchmark | "outperformed by 4%" | Name the index, the period, the currency |
| Unlabelled gross figure | Performance with no cost model | State the cost model or label it gross |
| In-sample presented as performance | A Sharpe with no IS/OOS marker | Mark it; if in-sample, say so in the same sentence |
| Survivorship in a cited aggregate | An industry-average return | Say whether the source is survivorship-conditioned |
| Paraphrase as quotation | A famous line without quote marks | Quote it exactly or mark the paraphrase |
| Wrong attribution | A concept credited to whoever popularised it | Check the primary source. (The "seven sins" are Deutsche Bank's, not de Prado's) |
| Contested claim stated as settled | VPIN described as predictive; HRP as dominant | Label the status and name the disputant |
The one question that catches most of it: could a competent reader check this claim, and if so, how? If the answer is no, the claim is decoration.
Pass 2 — Logic and mechanism
Mechanism check. For every causal claim, ask why the effect exists. A finance claim without a mechanism is a correlation with a story attached. If the draft says "the strategy works because it captures the premium", ask: which premium, paid by whom, in exchange for what risk?
Counterparty check. For every claimed edge, ask: whose losses are these? If the draft cannot name the counterparty's business model, it has not identified an edge.
The strongest-opposition test. Find the best argument against the draft's thesis. Is it in the piece? Is it answered? If the piece's opponent is a straw man, the piece has not been argued — it has been asserted.
Post-hoc mechanism check. Was the mechanism stated before the result was known, or constructed to fit it? If the latter, say so explicitly in the piece.
Correlation and causation. Flag every "drives", "causes", "leads to", "because of". Most should be "is associated with" or "is consistent with".
Base-rate check. A conditional probability stated without its base rate is misleading. "80% of funds that did X failed" is uninformative without how many funds did X and what the overall failure rate is.
Selection check. Was the evidence selected to support the thesis? Look for the pattern: three examples that all point one way, with no mention of counterexamples.
Pass 3 — Numerical verification
Recompute. Do not trust the draft, including your own.
- Arithmetic: re-derive every calculation. Percentages, annualizations, ratios, differences.
- Annualization: check the √q multiplier against the actual data frequency, and check whether autocorrelation was considered.
- Percent vs percentage point: a move from 2% to 3% is one percentage point, or a 50% increase. Check every instance.
- Units and currency: check for mixed units (basis points vs percent), mixed currencies, and nominal vs real.
- Period consistency: do all the compared numbers cover the same period? A strategy Sharpe and a benchmark Sharpe over different windows are not comparable.
- Rounding direction: does rounding flatter the author's thesis? Round both ways.
- Sum checks: do the parts add to the whole? Do attributed PnLs sum? (They should not, for overlapping risk classes — but the piece should say so.)
- Sign checks: a negative Sharpe written as positive; a drawdown written without its sign.
- Order of magnitude: does the number pass a smell test? A 400% annual turnover and a 400% daily turnover are different claims.
- Table vs prose: do the numbers in the narrative match the numbers in the table? A narrative that says "twenty standard deviations" over an appendix that says otherwise has published two contradictory documents.
Pass 4 — Register and structure
Register check. Identify which register the piece is in — practitioner-confessional, academic-polemical, journalistic-narrative, instructional (see quant-finance-writing). Then check for drift. Drift symptoms: a first-person anecdote in an academic argument; an equation in a narrative profile; a "Key Points" box in a polemic.
Claim position. Is the central claim in the first 150 words? If the reader must reach paragraph six to learn what the piece argues, move it.
Contract. Does the piece state what it will and will not deliver? If it does, does it keep the promise?
Section discipline. Does each section make one point, stated in its first sentence?
Paragraph length. Any paragraph over five sentences is a candidate for splitting.
Ending. Does the piece end on the unresolved question, or on a summary that resolves everything? The latter reads as marketing.
Format fit. Does the format match the evidence? A research note without a trial count should have been a teardown. An explainer carrying three concepts should have been three pieces.
Pass 5 — Craft line-edit
Only now do you edit sentences.
Cut, in order of yield:
- Intensifiers — very, extremely, hugely, dramatically, significantly (unless statistical). In a domain with numbers, an intensifier is a confession.
- Throat-clearing openings — "In today's markets…", "Since the dawn of…", "It is important to note that…"
- Nominalizations — "the implementation of the model resulted in" → "implementing the model"
- Adverb-hedge stacking — "could potentially possibly"
- Redundant pairs — "each and every", "first and foremost"
- Explaining what you are about to explain — "In this section, I will discuss…"
Strengthen:
- Replace abstractions with concrete nouns: "financing friction" → "borrow fee"; "universe attrition" → "delisted names".
- Land long sentences on short ones — and check that the short sentence adds a verdict the long one lacked.
- Put the number in the sentence that makes the claim, not in the next one.
- Convert passive to active where the actor matters.
- Check that every technical term is defined exactly once, on first use.
Sentence-level checks:
- No sentence over 45 words without a deliberate reason
- Every paragraph's first sentence is its topic sentence
- No more than two one-line paragraphs per 1,000 words
- Every analogy followed by where it breaks
- Every equation has its symbols defined immediately after
Pass 6 — Ethics and compliance
- Not framed as personalised advice; no present-tense recommendation on a named security
- Positions, employment, funding, data access disclosed — near the top if securities are named
- Standard past-performance disclaimer not used as a substitute for assessing an overfit result
- No MNPI; no individual-level PII republished
- Contested claims labelled; both sides given; disputant named
- Quotations verified; permissions respected
- Affiliations disclaimed precisely rather than with blanket boilerplate
- Corrections route stated
Full detail: finance-claims-and-ethics.
Red-flag phrase list
These phrases reliably signal weak finance writing. When you find one, look for the defect it is covering.
| Phrase | Usually conceals |
|---|---|
| "backtested and proven" | No out-of-sample test, no trial count |
| "consistently outperformed" | A selected period; no benchmark named |
| "mathematically sound" | Mathematical terminology without mathematical content |
| "the strategy has a Sharpe of X" (no interval, no N) | Estimation error and selection bias |
| "risk-adjusted returns" (no definition of risk) | Volatility used as a proxy for a tail risk it misses |
| "proprietary algorithm" | An unassessable claim |
| "past performance is not indicative of future results" (as the only caveat) | An overfit result being shielded |
| "AI-powered" | A model whose evidence is not reported |
| "we use machine learning to find patterns" | Trial count hidden |
| "in today's volatile markets" | Nothing |
| "smart money is buying" | No data |
| "the market is inefficient" (unqualified) | No statement of which friction, or how large |
| "there is no free lunch" | A refusal to engage with the specific claim |
| "nobody could have predicted" | A failure to state what was knowable |
| "the model is always learning" | No evidence of adaptation |
| "significantly better" (no test) | A difference with no significance test |
| "guaranteed", "can't lose", "risk-free" | Fraud or ignorance — stop and verify |
| "20 standard deviations" (unexplained) | A secondhand claim with no methodology |
The general rule: any phrase that asserts a quality instead of reporting a measurement is hiding something. Ask what the measurement would be, and whether it was made.
Self-review protocol
When reviewing your own draft, do these in order. The point is to break familiarity.
- Set it aside for at least a few hours. Same-day self-review finds typos and misses arguments.
- Read only the first sentences of each paragraph. Does the argument hold? This is the skeleton test.
- Read only the numbers. Do they tell a coherent story on their own? Do they contradict the prose?
- Read only the quotations. Do they support the claims they are attached to?
- Read backwards, last paragraph first. Catches continuity errors and reveals whether the ending depends on the introduction.
- Read aloud. Every place you stumble is a place a reader stops.
- Cut 10%. Then decide what to restore. Restoration is a decision; cutting is a reflex.
- Ask the adversary question: if someone hostile read this, where would they attack? Fix that place or acknowledge it.
Review output format
When reviewing someone else's draft, structure the response:
- Verdict in one sentence. Is the argument sound? Is it supported? Should it be published?
- The three highest-value fixes, ranked by impact. Not a list of everything.
- Evidence failures, enumerated with the claim quoted and the defect named.
- Logic failures, with the specific unstated assumption identified.
- Craft notes, grouped by pattern rather than by instance (do not list twenty comma splices).
- What works, specifically. Vague praise is useless; "the mechanism-first ordering in section 3 is doing real work" is useful.
Do not rewrite the piece. Name the defect and let the author fix it. A reviewer who rewrites has taken ownership and the author has learned nothing.
Writing guidance
Do
- Review against the piece's own contract, not against the piece you would have written.
- Rank fixes. A review with thirty equally-weighted notes is a review with no priorities.
- Name patterns, not instances. "You hedge every quantitative claim twice" is actionable; fifteen examples of the same hedge is not.
- Praise specifically. Say which sentence is doing the work.
- Separate wrong from unfashionable. The first is a defect; the second is taste.
- Check the numbers yourself before commenting on style.
Don't
- Don't line-edit before the evidence audit.
- Don't confuse confidence with quality. A confident piece can be wrong; a hedged piece can be right.
- Don't demand the standard past-performance disclaimer as a substitute for the trial count.
- Don't let a well-written paragraph escape the evidence check because it reads well. Fluent prose is where unsourced numbers hide.
- Don't ask for a mechanism the piece cannot have — a narrative profile does not need one, a research note does.
- Don't review the draft you wanted to read.
A reusable framing: Read for the argument, then for the evidence, then for the numbers, and only then for the sentences. Fluency is the last thing to check, because it is the thing most likely to hide everything else.
See also
quant-finance-writing— the standards this rubric enforcesfinance-claims-and-ethics— the ethics pass in fullresearch-integrity— fraud, hidden trials, misconduct claimsfinance-article-formats— format fit, and the evidence each format requiresfinance-narrative-craft— craft criteria for narrative piecesexplaining-quant-concepts— the analogy and worked-number checksbacktest-pitfalls— the reader's audit for backtest claims