I. The Loss Is Not Random

A building in Austin generates income. Over twelve months: tenants pay rent, some late, some early. One renews after their competitor closes. The property manager handles a leak that takes three calls. The owner covers an unexpected assessment. Eleven months of mundane reliability and one month of genuine uncertainty.

All of this becomes a single row: Net Operating Income, $847,000.

This is compression. Capital markets exist because of it — you cannot compare two buildings without standardizing their representations. But look at what the $847,000 discarded. The tenant relationship that drove the renewal. The property manager’s judgment about the leak. The competitor closure that changed the probability. The number kept what could be verified and ate everything else.

That tradeoff is usually worth it. A REIT investor with 200 properties needs comparable numbers — nobody wants 200 narrative accounts of tenant relationships. The $847,000 is good enough for the portfolio investor. For the single-asset buyer doing direct due diligence, it fails them completely.

That loss is also directional. Financial compression is biased toward auditability and away from what would help you predict. For example, the lease term is in the rent roll because it’s verifiable. The tenant’s likelihood of renewal is not, because it’s judgment. For the investor, the assessment of renewal probability is probably more useful.

Even the verifiable part has discretion baked in. Two accountants can produce different NOIs from the same building without either of them being wrong (e.g. depreciation method, expense classification, revenue timing, and so on).

The loss is also invisible. A zip file knows it’s compressed. The $847,000 arrives with no metadata about what was included or excluded. The format has no way to describe what it left out.

II. Claims, Not Measurements

A thermometer measures temperature. If it reads 72 and the room is 68, the thermometer is wrong.

Financial numbers don’t work this way. When a company reports $2.3 billion in quarterly revenue, that number passed through recognition policies, cutoff decisions, and judgment calls about contingencies — all made by people with interests, under time pressure, following rules that permit discretion.

A 10-K is produced under real constraints — GAAP, SOX controls, audit standards, restatement risk. These constraints work. But the 10-K still has to satisfy SEC requirements while minimizing legal liability, and that leaves room. Authors make choices within the rules, and those choices tend toward what survives audit over what would actually help you predict.

III. Model the Compressor

What does it mean to be good at reading financial numbers?

The obvious answer is technical skill — better models, deeper accounting knowledge. But technical skill operates on the numbers as given. It cannot recover what the compression discarded.

A different skill matters more: modeling the compressor. If every number is a claim by an author optimizing for audit survival, understanding the number requires understanding the author. What pressures were they under? Where did they have discretion? What would they have wanted to emphasize?

A 10-K is evidence of what happened and evidence of how the author wanted the reader to see what happened. So channel checks and management meetings matter even though they look nothing like analysis. They provide information about the compression process itself — metadata the format can’t carry.

In practice: a company reports flat revenue but cash collections are declining and receivables are growing. The number says stable. The compression hid a deterioration. Why would management want to show flat revenue? What recognition choices would produce that? What does the cash flow statement say that the income statement doesn’t? You learn to read the gaps between the numbers the author chose to show you.

This skill is rarer than it should be outside of wallstreet. Years of being rewarded for reading numbers one way trains you to miss what the numbers can’t say. The engineer who has processed numerical datasets reaches for that frame when encountering finance. Soft information is filtered out, not from laziness but because the trained frame doesn’t flag it.

And institutional incentives reinforce this as any documented uncertainty becomes a litigation risk. “I considered 11% but chose 10.5%” would be a liability - so the rational analyst hides the range.

IV. Reference Numbers

Some financial numbers become the reference points that everyone uses. A cap rate, a credit rating, a stock price — people build contracts around them, make decisions off them, treat them as shared facts.

The problem is nobody audits the inputs with the same intensity they trust the output. The reference number looks clean. The claims that built it were not. The longer the chain between raw reality and the reference number, the more compression and flattery accumulates — and the less anyone can trace what went in.

That’s 2008. Originators chose and paid the agency rating their loans. Buyers did almost no loan-level due diligence. Correlation assumptions in CDO models were untested. Every link in the chain had misaligned incentives, and the compression made it invisible. A functioning index fund is also compression — but without the misalignment at every handoff. In 2008 you had both, feeding a rating that synchronized trillions. When it broke, everyone broke together.

Most numbers you trust were produced under constraints that don’t prioritize your needs. The precision is the problem. It can make you forget there was ever a range.