Could You Reconstruct Why You Sold in 2021? How to Document Investment Decisions So the Answer Survives

To document investment decisions properly you need five things on the record, not one: the date and instrument, the rule that fired with its threshold as written that day, the state of your book at that moment, the action you actually took, and the version of your policy that was in force. Most people write down a reason. A reason alone does not survive five years, because it does not carry the threshold that made it a reason.

Here is the test, run below on an illustrative book. On 12 March 2021 a sale goes through: 400 shares of a long-held position at $367.00, for $146,800. Five years on, could you say which rule produced that number? Not the story. The rule.

What does it mean to document investment decisions?

Search the phrase and you get the same advice everywhere: keep an investment journal, write down your thesis, review it later, learn from your mistakes. It is decent advice and I have no quarrel with it. It is also not sufficient, and the reason is specific rather than vague.

A journal records what you believed. An audit trail records what you were operating under. Those are different objects, and only the second one can be checked. If I write "trimmed the position, margin felt high," I have preserved a feeling. Five years on that sentence is untestable, because I no longer know what "high" meant to me in March 2021, and I no longer have the book it referred to.

I run a leveraged, multi-strategy book myself, and this is not a hypothetical failure for me. When I went looking through my own records for the reasoning behind a set of 2021 trades, I found the confirmations, the prices, and the realised gains. What I could not find was the rule. I knew what I did. I could not prove why. That is the gap this post is about — and the worked example below is a constructed one, built to a realistic shape so the arithmetic is followable, not a printout of my account.

This matters more than it sounds, because an unreconstructable decision is one you are liable to make again without knowing it. That is the governance problem underneath the record-keeping problem, and it is why this sits in portfolio governance rather than in productivity advice.

The reconstruction: one sale, five years later

Here is the state of the illustrative book on 12 March 2021, the day of the sale.

The book on 12 March 2021 (illustrative figures)	Amount
Gross securities position value (holdings at market)	$2,318,000
Margin loan (the account's negative cash balance)	$811,500
Margin utilization (loan ÷ securities)	35.01% — Freeze

Two things have to be said about that last row before it can carry any weight, and they are the whole thesis of this post in miniature.

First, the denominator. That 35.01% is margin loan divided by the gross market value of the securities, before the loan is subtracted. Our convention, not an industry standard. Run the identical book on a net-of-loan denominator — loan against account value after the loan comes off — and $811,500 against a net $1,506,500 is 53.87%. Same book, same day, same loan; one number sits mid-range and the other looks alarming. Neither is wrong — they answer different questions. That is precisely why a threshold is meaningless unless its definition travels beside it, and it is why "margin was too high" fails as a record.

Notice what is not in that table: a cash line. A book carrying a margin loan does not have spare cash sitting beside it — the loan is the negative cash balance. If $87,000 were sitting in the account, it would already be inside the debit, and the loan would read $87,000 smaller. The only ways to show both at once are a positive balance in a different currency, or cash the broker genuinely has not swept against the debit. Neither is the normal case, and a worked example that quietly assumes one is a worked example you cannot re-run.

Second, what "Freeze" means and what it does not. These are four self-imposed bands — Clear below 30%, Harvest from 30% up to but not including 35%, Freeze from 35% up to but not including 40%, Forced 40% and above — and 35.01% lands just inside Freeze. The zones are introduced on margin as a weekly management tool, and the boundary convention that puts 35.01% in Freeze rather than Harvest is set out in the difference between a dashboard and a governance instrument.

The sale itself: 400 shares at $367.00, for $146,800, with the proceeds applied against the loan. That leaves a loan of $664,700 against securities of $2,171,200 — 30.61%, which is Harvest. So the sale moved the book down one zone. That is the entire factual record. Now the question: which rule told you to do it?

Which rule fired, and what the arithmetic cannot settle

Three written rules could plausibly have produced a sale that day. Each implies a specific size — and that is the lever, because a rule does not merely tell you to act, it tells you how much. Size is a fingerprint. (These three are illustrative policies I have invented for the exercise, not published Incomestead settings.)

Candidate one — margin, reduce out of Freeze. The book was at 35.01%, barely over. Bringing it below 35% takes $307.69. But a reduction has to be rounded up, away from the band edge, to a stated increment, and then re-checked at the precision you actually display: $400 computes to 34.9974%, which shows as 35.00% at two decimals and so still reads as the band you were trying to leave. Say the rule rounds up to the next $5,000. Then it instructs $5,000 — 34.87%.

Candidate two — margin, reduce all the way to Clear. Getting under 30% takes $165,857.14 exactly; $166,000 computes to 29.9954% and displays as 30.00%, so it fails the same test. Rounded up to the next $5,000, the rule instructs $170,000 — 29.86%, leaving $2,900 of room before the reading would touch 30% again.

Candidate three — allocation, rebalance an overweight Layer. A Layer — one strategy inside the Incomestead Stack — is the tier sitting between your Buckets and the individual portfolios inside them. Say the Income Engine Layer held $603,000 — 26.01% of the book against a 22% target with a ±3 band, so outside its 25% upper bound and genuinely in breach. (Our real bands are proportional rather than fixed points; this one is invented for the exercise.) Sending it back to the midpoint of target and the breached edge, (22+25)/2 = 23.5%, takes $76,169.93, which on the same $5,000 increment instructs $80,000.

That third number contains a trap worth pausing on. It is $76,169.93 only because the proceeds pay down the loan, which shrinks the book and therefore moves the target the sale is chasing. If instead the proceeds are redeployed rather than repaid, the same rule instructs $58,270 — 23.5% less for an identically worded rule. A rule that does not say where the money goes is not a complete rule, and you cannot reconstruct which version you were running from the sentence alone.

The increment is the same kind of trap, and I have to concede it against my own table. Had the rule said "round up to the next $1,000" instead of $5,000, candidate one would instruct $1,000, and the sale below would read as 146.8× the instructed size rather than 29.4×. Same book, same rule, same reasoning — and a fingerprint that moves by a factor of five on an increment nobody wrote down. Size only identifies a rule if the rounding is part of the rule.

If the rule was…	It instructs	Against the $146,800 actually sold
Margin — reduce out of Freeze	$5,000	Sale is 29.4× the instructed size
Margin — reduce to Clear	$170,000	Sale falls $23,200 short
Allocation — rebalance the Layer (proceeds to the loan, $5,000 increment)	$80,000	Sale is 1.84× the instructed size
Actually sold	$146,800	Matches none of the three

Now the honest part, and it is the reason this exercise is worth your time rather than mine. That table does not show the sale was discretionary. It shows something narrower and more useful: none of three dollar-denominated rules produces that size. At least six explanations survive, and I cannot separate them with the numbers I have.

Start with the possibility that it was not a decision at all. Four hundred shares is four option contracts, and a sale at a flat $367.00 is consistent with a covered call being assigned — the stock leaving at a price agreed weeks earlier, no rule firing and nobody choosing anything. I cannot prove that from the record either: a limit order resting at $367.00 prints the same line, and so does a round-lot sale that happens to fill at a whole dollar. That is the whole problem in miniature. But if it was an assignment, the only decision worth recording happened weeks before the sale, when the call was written. A record that assumes every sale was a decision files this one under a reason that never existed.

The rest. Any rule expressed in share counts ("trim 400," "sell down to 600," "cut the lot by a third") produces this same signature, and a dollar-fingerprint test is structurally blind to it. A position-level rule — a single-name concentration cap, a trailing stop — cannot be tested at all, because the book state above never records how large that position was. Two rules may have fired together. The proceeds may not all have gone to the loan, which would break every reconstruction above, since each one assumes they did. And a sale can be forced from outside the book entirely, by a tax bill or a capital call, which is not a governance decision in any sense.

There is one more gap, and it is the largest. Under a governed process a breach does not act on a single reading: the two-week persistence rule requires a breach to persist across two consecutive weekly snapshots before it drives an action. (40% and above — the Forced zone — is the one exception, acted on the week it is first read; this book is not there.) Which means that on one day's book state, none of these three rules necessarily instructs anything. To know whether the margin rule was even live on 12 March 2021 you need the reading from the week before — and that number appears nowhere in the record. A rule that fires on persistence needs two book states, not one. Worse, the reading a weekly rule acts on is the snapshot, not the trade-day book, and nothing above establishes that the 12 March state is the one the rule actually saw.

Why writing down the reason is not enough

Suppose you had been diligent. Suppose that in March 2021 you wrote, in a notebook, "Sold to bring margin down — utilization was too high." You have now done what every article on the subject tells you to do. Has it worked?

No, and here is precisely why. Take a reading of 33.2%. Under a personal ceiling of 30% that is a breach, and the instruction is quantified: reduce by a specific amount. Under the four bands it is Harvest, where the instruction is qualitative — trim, raise cash, let income pay the line down, while it is still your choice rather than your broker's. Both regimes point the same direction, but only one tells you how much, and the note "utilization was too high" is equally consistent with both. It cannot tell you which regime produced the size you actually traded — and size, as the table above shows, is the clearest fingerprint you get.

This is the load-bearing idea, and the generic advice rarely says it: a decision record has to store the policy version that was in force, not just the reason. A reason is a sentence. A policy version is a set of numbers with a date on it. Only the second lets you take a decision from five years ago and re-run it — feed the old book into the old rules and see whether the output matches what you actually did.

It also means each week has to be judged under the rules of its own era. If you tighten your margin ceiling in 2027, that does not retroactively make your 2021 decisions violations. Re-running history through today's policy manufactures false guilt and false comfort in roughly equal measure. The old week gets judged by the old rules — that is what makes the judgement worth anything.

There is reasonable evidence that writing the rule down in advance is where the benefit sits. Gollwitzer and Sheeran's 2006 meta-analysis of implementation intentions — pre-committed "if-then" plans — pooled 94 independent tests from 63 reports and found a medium-to-large effect (d = .65) on goal attainment versus holding an intention in your head (Advances in Experimental Social Psychology 38, 69–119). Two caveats have to travel with that number, and they cut hard. A much larger 2024 meta-analysis across 642 tests puts the effect at d = .36, and as low as d = .15 under a Bayesian correction — so the 2006 figure is very likely inflated. And the underlying studies are overwhelmingly health and academic behaviours, not investment decisions, so applying any of it here is an extrapolation. The mechanism is what transfers — you decide once, in the calm, so the situation triggers the action instead of your judgement relitigating it in the moment.

On the cost of ungoverned timing decisions, the honest position is that the size is disputed, and I would rather show you the dispute than pick the flattering number. Morningstar's Mind the Gap 2025 study puts the shortfall between investors' dollar-weighted returns and their own funds' total returns at about 1.2% a year over 2015–2024, attributing it to the timing of purchases and sales — while cautioning that the finding should not be read as a parable about dumb money. That estimate has been directly challenged: Fulkerson, Jordan, Riley and Yan, in Bad Timing Does Not Cost Investors 15% of Their Funds' Returns (Financial Analysts Journal, 2026), find that timing accounts for only about 0.10% a year. (Morningstar's figure is 1.2 percentage points — 7.0% against 8.2% — and reading the difference as "real but largely misattributed" is my gloss on the dispute, not their wording.) Take whichever you find persuasive. Neither touches the point I am making: whatever timing decisions cost in aggregate, you have no way to audit your own if you never recorded them.

The Five-Field Decision Record

Most of the fields below exist because the reconstruction above failed without them — they are the list of things that were missing, not a template I decreed.

Run it this week, on your own numbers. Not on a hypothetical, and not on the next trade you make — on the most recent sale you actually placed. Open the confirmation and try to fill in all five fields.

If you can complete fields 1 and 4 but stall on 3 or 5, that stall is the finding. It means a decision you made recently is already unreconstructable, and it has only been a few months. Keeping this record on a cadence rather than when you remember is its own discipline, covered in why the weeks you skip are the dangerous ones. Whether a spreadsheet is a fit home for it is the subject of what the spreadsheet version actually costs you — the short answer being that a sheet will hold these five fields perfectly well and simply will not stop you leaving them blank.

What a decision record will not do for you

I want to be straight about the limits, because a record is a modest instrument and it is easy to oversell.

It will not make your decisions correct. A well-documented bad decision is still a bad decision; what you gain is the ability to identify it as one later rather than remember it as reasonable. It will not tell you what to do next — the record is evidence, not instruction. It will not protect you inside a fast market, as the drawdown figures above should make plain. And it will not substitute for writing the policy in the first place: a record of decisions made under no policy is a record of improvisation, tidily filed.

What it does do is move your process out of your memory — which is the single point of failure in almost every solo operation, mine included — into something that exists outside you, can be checked, and can be handed to someone else. A written, dated policy is also one of the four components of the Governance Score, the single number we use for how well a book is being governed; the decision record is what proves you actually operated the policy rather than merely owning one. And the Score's margin gate bites regardless of the paperwork: a book sitting in Freeze, like the one above, is capped at 65 no matter how good its records are.

Frequently asked questions

How do I document investment decisions without it becoming a second job?
Five fields per closed decision, not per thought. In most weeks you will make no entries at all, because in most weeks a governed book does not trade. The record grows at the rate your decisions do.

What is the difference between an investment journal and an audit trail?
A journal stores what you believed; an audit trail stores what you were operating under. The practical test is whether a past decision can be re-run: feed the book state from that day into the policy version in force that day and see whether the output matches the action you took. A journal as usually kept fails that test, because it records reasons rather than thresholds. Kept for a successor rather than for yourself, this is the forward version of the same record problem.

Why does the policy version matter more than the reason?
Because a reason without its threshold is ambiguous. A margin utilization reading of 33.2% is a quantified breach under a personal 30% ceiling and a qualitative "trim and raise cash" state under the four bands — the note "utilization was too high" is consistent with both, so it cannot tell you which one produced the size you traded.

What is a Layer?
A Layer is one strategy inside the Incomestead Stack — Bedrock, Pillars, Long-Horizon Growth, Short-Horizon Growth, Income Engine, Precious Metals, Crypto or Cash. A Bucket is the tier above it, and each Layer holds one or more portfolios you name yourself. An institution would call a Layer a "sleeve"; we use Layer.

Should I go back and reconstruct old decisions?
Reconstruct the most recent one as a diagnostic, because the exercise shows you how much is already unrecoverable. Beyond that the effort rises and the yield falls quickly — and reconstruct old weeks under the rules that were in force then, not the rules you hold today, or you will manufacture violations that never happened.

This is education, not personalised financial advice. I am a practitioner writing about how I govern my own book, not a licensed adviser, and nothing here is a recommendation to buy, sell, or borrow. I am also the founder of Incomestead, which sells software that does what this post describes — weigh the argument accordingly. Margin amplifies losses as well as gains, and a broker can liquidate positions without contacting you first. All figures in the worked example are illustrative.

Incomestead recommends. You decide.