Finance Is the Feedback Loop

The Smoothing #

Forecasts get smoothed. I've participated in that smoothing. The practice remains different from inventing numbers, even though the word sounds...suboptimal. The distinction matters. It is, more or less, the entire point.

The quarter had come in uneven, which is a polite way of describing what happens when sales bookings land in unpredictable lumps and the revenue recognition schedule you built in October starts to look like a document written by someone who lived in a more orderly universe, three universes over. Sales believed the gap would close, which is what Sales usually believes, because Sales operates on a fundamentally different theory of time than Finance does.^[1]

Our particular variance line looked volatile in a way that would invite questions we didn't yet have a coherent explanation for, and if you've ever presented a volatile variance line without clean answers to someone senior, you know that the subsequent forty-five minutes of your life will be consumed by a kind of Socratic interrogation that produces some insight but much more anxiety, after which someone will recommend that you "get closer to the drivers," which is the governance equivalent of a doctor telling you to "be healthier."

The instinct to make the line easier to explain can sound evasive from the outside. Sometimes it is. More often, the problem is that volatility doesn't remain contained to the line where it appeared. It moves. The board deck turns it into a question whose answer may alter hiring. The executive conversation turns it into a decision that now feels exposed. Downstream, teams start asking what they are supposed to plan against if the number itself may move again next week.

I've watched organizations react to every new piece of information as though it were a final result. A soft bookings quarter slows hiring, followed by a strong quarter that restarts it, followed by a slipped deal that brings the whole discussion back. People begin rebuilding plans around movement rather than direction. Eventually nobody truly believes the forecast, although everyone still has to use it.

At some point, almost without anyone saying it out loud, the number stops functioning solely as a measurement. It becomes something closer to a coordination surface. Product plans, hiring decisions, sales targets, and cash expectations all rest on it, and coordination surfaces only work if people trust that they will hold still long enough to build on.

So we adjusted the weighting and tempered the swing. Smoothed. Smoothing can make interpretation easier, and it can keep an organization from mistaking ordinary timing for a change in the business, but it can also delay detection by narrowing the amplitude of a signal that deserved to remain visible.

That's the slippery part. In the moment, noise being filtered and signal being damped can look almost identical. A January bookings shortfall might be random variance or the first indication of a market shift. A March spike might confirm that the quarter was merely uneven, or it might be pull-forward borrowing from April. You don't know yet. You can't know yet, and the entire organization would still prefer a number by Friday.

Leaving every fluctuation untouched does not solve this. Raw volatility carries its own fiction, the suggestion that the latest movement deserves more weight simply because it arrived most recently. The organization becomes sensitive to everything and therefore capable of distinguishing very little. Smoothing too aggressively creates the opposite problem. A structural shift keeps getting classified as timing until the trend becomes obvious enough that the useful window for acting on it has narrowed. Both choices can feel correct and responsible at the time you are making them, which is how most difficult decisions manage to remain difficult.^[2]

Assumptions are arguable. They are assumed. Once you start modeling scenarios instead of reporting raw deltas, you're operating inside a possible world rather than merely describing the one that has already occurred. So you document the assumptions, make them visible, and keep the outputs within a defensible range. Those practices don't remove judgment. They leave enough of a trail for someone else to inspect how the judgment was made.

This is usually the point where the clean version of the job stops lining up with what it actually feels like to do it. On paper, the process is straightforward: measure, compare, adjust, improve. Numbers in, insight out.

In practice, by the time the numbers reach you, they have already been shaped. Not falsified, and usually not manipulated, but: nudged, weighted, framed. Someone decided what counted as pipeline. Another someone chose the probability applied to it. Someone classified a delay as timing rather than deterioration. Yet another person included a customer conversation that hadn't yet appeared in the system because leaving it out would have made the forecast technically cleaner and plainly less accurate. By the time a figure makes it onto a slide, it has passed through judgment.^[3]

We used to run a monthly forecasting competition at one company I worked at. The rules were almost comically simple: five high-level numbers, forecast for the upcoming month. Revenue, expenses, cash, headcount, and one or two others depending on the quarter.

Five numbers. No segment-level detail or unit economics or probability-weighted scenario analysis with seventeen tabs and a color scheme inherited from someone who left three reorganizations ago. Just five large numbers. In theory, the team should have been able to predict with reasonable accuracy given that we spent approximately all of our waking hours immersed in the data that produced them.

We could get close. Collectively, the team usually landed within a useful band. Someone was always off, though, and never consistently the same person or the same number. The person who nailed revenue would whiff on cash. The person who called expenses would miss headcount because a late-month hire start date slipped by a day and crossed the period boundary.

Five numbers. A team whose professional existence revolved around understanding them. Still, the exercise produced an uncertainty that humbled everyone who participated.

What made the competition useful was that nobody could dismiss the misses as a competence problem. The people making the forecasts knew the business. They also knew different parts of it, carried different conversations in their heads, and gave different weight to information that hadn't yet become data.

One person assumed a renewal would close on the 28th because it always had. Someone else knew the customer had postponed an internal meeting and expected the signature to slip into the next month. Both forecasts were reasonable. One would still turn out to be entirely wrong.

The exercise made visible something that polished forecasts tend to hide. "Knowing the business" is always partial. The confident voice presenting the number may understand the model better than anyone else in the room, but the model has been assembled from information distributed across people who each see a different edge of the same thing.

A forecast has to turn those partial views into a coherent account because the organization needs something it can act on. The account also has to remain open to revision, because coherence can become dangerous when everyone forgets how much uncertainty was compressed to produce it.

The uncomfortable question is how much clarity we can create without losing the signal that should change our minds.

The Loop #

Finance often gets cast as the department of rejection slips and red ink. The people who say no. The late arrival at the meeting who ruins the fun by pointing out that the plan is too expensive.

That framing misses something fundamental about the architecture. Finance isn't outside the system, evaluating it, but moving through it.

Accounting keeps the historical record. FP&A simulates possible futures. Strategic finance ties resources to bets. Systems and data connect decisions to what happened afterward.

Together they form a feedback loop that, in its healthiest version, works roughly like this: you design a plan, allocate resources to it, surface what happened against what you expected, interpret the gap between plan and reality, and update the next cycle accordingly.

When the loop runs, decisions compound into learning. Decision A leads to outcome C, the system remembers enough about how it got there, and the next time the organization faces a similar choice it begins with accumulated evidence rather than accumulated confidence, which are extremely different things.

Every stage still contains judgment. A plan expresses a belief about the future. Execution changes the conditions the plan was trying to anticipate. Detection depends on definitions, systems, and timing. Analysis requires someone to distinguish noise from changed circumstances. Adjustment requires the organization to respond without rebuilding itself around every miss.

The loop works through correction rather than certainty. A forecast is a working theory made precise enough to act on. The variance is reality returning with notes.

When the loop stalls, the work rarely stops. Reports circulate. Plans get refreshed. Dashboards glow. Meetings continue with the usual impressive volume of activity, which is exactly why the loss of learning can take so long to notice.

Where It Breaks #

Before you can learn anything from the numbers, you have to structure the data that produces them. Structuring data is roughly 60 to 95 percent of the job, depending on the month, and almost none of the part anyone means when they talk about "the job."

I mean this literally. The elegant feedback loop assumes that you can connect a decision to its outcome, that when you fund a bet you can later trace the spending to the result and understand what happened. In theory, this is straightforward. In practice, it requires chart-of-accounts design, cost-center hierarchies, tagging conventions, allocation methodologies, and data-pipeline maintenance that can be so tedious to build and so fragile to maintain that most organizations do some portion of it poorly and then wonder why the resulting analysis feels thin.

You cannot learn from data you can't structure and keep structured. You can't structure data that lives in fourteen systems with inconsistent naming conventions and no shared identifier. The time required to solve these problems is time nobody outside the connective functions tends to see, because the output is plumbing, and plumbing remains invisible until two teams present contradictory customer metrics in the same meeting and everyone discovers that one system counts parent accounts while the other counts billing entities.^[5]

Even the obvious answer, fix the plumbing, contains a tradeoff. Every hour spent reconciling definitions is an hour unavailable for the analysis the business already needs. Waiting for perfect structure may produce a cleaner answer after the decision has passed. Moving ahead with imperfect data may preserve the moment and embed an error nobody notices until later.

The loop breaks elsewhere when Finance receives a plan after the bet has already been socially committed.

More than once, I've reviewed assumptions after an initiative had been announced internally, teams had aligned around it, and people had begun arranging their work as if approval were a formality. I identified the gaps and documented risks. Sometimes I pushed. Sometimes I moved on.^[6]

"Pick your battles" sounds like advice about courage, but it is usually advice about proportion. Every challenge spends time, trust, and some portion of your credibility. Asking the room to reopen a committed plan may prevent a larger mistake. It may also create weeks of disruption around a risk that is real but tolerable.

The social mechanics matter because Finance can only pull the organization backward so many times before people start finding ways to move without it. Push on everything and review becomes obstruction. Push on nothing and analysis becomes documentation written after the decision. Knowing which concern deserves escalation would be much easier if the consequences arrived before the choice.

Variance creates another break in the loop because it produces an immediate demand for explanation. The room needs to know whether the result changes the plan. Teams need to know what they are supposed to do next. Nobody can remain suspended forever inside the intellectually pure position that more analysis is required.

Explanation serves a purpose. It gives the organization a provisional account it can coordinate around.

The trouble begins when the provisional account hardens before it has been examined. "Timing" may be exactly right. So may "one-time item," "deal slippage," or "temporary softness." Each phrase can also become a drawer where an unresolved variance gets placed so the meeting can continue.

Curiosity starts with what we misunderstood about the system. Explanation starts with the story that lets the existing model survive the result. Healthy analysis needs both. Curiosity without a stopping point can become a refusal to decide. Explanation without a return visit can prevent the loop from closing.

I've done both, and neither failure is the one I expected to have. Explanation is easier, faster, and far more rewarding in the short term because it lets the room move on, and sometimes the room genuinely needs to move, while other times momentum becomes the mechanism by which the same assumption survives one quarter too long. The opposite mistake is quieter and, for me, more habitual. I've stayed with variances long after the room had extracted everything useful from them, partly because the unanswered question bothered me and partly because spreadsheets are more comfortable than decisions. The version of that story where I am the last honest person in the room is available and I've told it to myself, but the more accurate version is that a spreadsheet can't be disappointed in you and a room full of people waiting on a recommendation can.

Dashboards live inside a similar tension.

At one company, I spent three weeks building dashboards that were, by any technical standard, excellent: clean visualizations, real-time data, appropriate drill-down capability. I sent them to distribution lists of thirty or forty people every week, on schedule, and I was genuinely proud of them.

People opened the dashboards before meetings. They used them to prepare. That mattered. Shared numbers made the meetings faster and reduced the number of arguments that began with two people reading from different files.

Six months later, during a strategic review, the head of product asked a question about customer payback period that the dashboard answered on its second tab. Nobody referenced it. Nobody had opened that tab in weeks.

The dashboard worked in the narrow sense that the data was there, refreshed, and available to anyone who wanted the answer. What I had built was access. I'd apparently assumed, without ever saying so, that access would produce attention, attention would produce examination, and examination would become learning. None of those things follow automatically from the one before.^[7]

The instrument was useful. We'd never built the rhythm around it that would make people return after a decision and ask whether the result matched the original belief.

A dashboard becomes part of the feedback loop when it helps change the next choice. Until then, it is infrastructure, possibly very good infrastructure, waiting for an organizational habit that may or may not arrive.

Forecasts can stop at the same point. They get polished until they look coherent enough to coordinate around, while the uncertainty removed in the polishing becomes difficult to recover. Which returns us to the smoothing.

Oscillation #

Feedback loops can and do sharpen judgment. They can also overcorrect, and the overcorrection pattern in finance is consistent enough across organizations that it probably deserves its own taxonomy.

After a painful miss, Finance tightens. The response is usually fast and comprehensive. Hiring slows. Discretionary spend requires additional approval. Forecast assumptions get stress-tested with a rigor that would have been useful six months earlier but is now being applied retrospectively to a plan that has already failed. Experiments shrink. Risk tolerance contracts. The system protects itself.

Those changes often work. Cash lasts longer. Commitments become more visible. Plans rely less heavily on outcomes that have not arrived yet. A system that had grown loose regains some discipline, and the company buys itself enough room to find out whether the miss was temporary or whether something more fundamental has changed.

The costs arrive elsewhere and at different speeds. A paused hire had already been embedded in a roadmap, a revenue target, and the private calculations current employees made about how long they could keep covering work meant for someone else. A delayed experiment preserves cash while postponing the answer the experiment was designed to produce. An additional approval catches weak spending and teaches everyone that smaller ideas are easier to move through the system, so signal may improve while ambition contracts alongside it.^[8]

The opposite failure is less visible at first. A company decides the miss was timing, keeps every hire open, preserves every experiment, and carries the old plan forward because changing it would create disruption of its own. Sometimes the quarter comes back and everyone feels vindicated. Sometimes another miss arrives before the first one has been understood, and the eventual correction has to be much harsher because the smaller one was postponed.

This is why tightening feels rational even when everyone in the room understands what it may do to the organization. One response protects possibility by preserving resources. The other protects possibility by preserving motion.

Finance sits at the junction of those two forms of belief, which gives it enormous power to shape what the organization attempts. Approval tells a team that an idea deserves resources. Delay communicates doubt even when the stated reason is procedural. A tighter forecast can prevent a bad commitment while quietly narrowing the range of ideas people are willing to propose.

Calibration matters more than permanent control. The system has to become more sensitive when conditions deteriorate, then relax when the evidence supports it, and the second movement is harder than the first. Adding a control has a visible reason. Something happened, and the organization responded. Removing the control asks someone to accept the possibility that the original problem may return. Safety gathers defenders. Friction accumulates more quietly.

Temporary responses can become the normal range of motion without anyone deciding that they should. Teams adapt to smaller budgets, slower approvals, and vacancies that were initially described as temporary. Plans begin arriving already shaped to fit the constraints. Eventually the organization may appear disciplined partly because it has stopped proposing the things most likely to challenge the system, and nobody has to want a smaller company for the company to become smaller in what it is willing to attempt.

Good calibration requires remembering why the system changed, watching whether that reason still applies, and accepting that the absence of another bad outcome does not prove every protective measure remains necessary.

When It Works #

The healthiest version of the loop is quieter than people expect.

A model cell turns red early enough that the room leans forward with curiosity instead of bracing for impact. Someone says "the variance is interesting" and means it, genuinely finds it interesting rather than threatening, because the cash is fine and nobody needs an immediate culprit.

You forecasted X and got Y. The assumption may have been wrong. The assumption may have been reasonable and something else changed. The miss may reveal a dependency the model never captured or a sensitivity that had been underweighted. Timing may explain the whole thing, although the team has enough discipline to establish that rather than reaching for the word because it is available.

These are good questions. Pursued honestly, they make the next forecast better and the next bet smarter.

I remember one review where a team came in below plan for the second month in a row. The first miss had been explained as timing, reasonably enough. The second made the explanation less comfortable. Nobody was in trouble. The overall quarter was intact. We could stay with the question without turning it into a trial.

We pulled the segment apart and found that the model assumed new customers would ramp at roughly the same rate as the prior year's cohort. They weren't. The customers were still arriving on schedule, which is why the miss had been easy to read as timing, but implementation had slowed and the revenue was following them later than expected.

The result did not produce a dramatic reversal. We changed the ramp assumption, moved a portion of the quarter into the next one, and adjusted hiring slightly before the gap became a cash problem. The next forecast was still wrong, but wrong by less and for different reasons.

That is what the loop looks like when it works. A miss changes the model. The model changes the plan. The next result returns with new information.

The room still needs an answer before every uncertainty has resolved. Curiosity cannot become an endless refusal to decide. A useful explanation can remain provisional, support action, and still carry a date when someone will return to see whether it held.

That return matters. Plenty of organizations inspect a miss, agree on a plausible story, and record an action item. Fewer come back later and test the explanation itself.

Our forecasting competition worked because the stakes were a Slack emoji and some gentle ribbing rather than a performance review. People exposed their real assumptions because being wrong carried very little cost.^[9]

Higher-stakes rooms need more deliberate care. They have to distinguish the quality of a judgment from whether the eventual outcome cooperated. They have to ask what was knowable at the time, whether the assumptions were visible, and whether new evidence changed them quickly enough. The point is to leave both the smoothing and the explanation visible enough that someone can inspect them, test them, and revise them when the evidence moves.

Finance doesn't earn trust by blocking every bad idea. You can block ideas all day and the organization will eventually learn that the easier path runs around you.^[10] Trust comes from helping the company see what happened clearly enough to choose what happens next with something approaching real understanding rather than sophisticated guessing.

The forecast will still miss. The dashboard will still lag. Information will remain partial, and some signals will arrive after the useful moment has passed. A working system keeps those limits visible inside the tools meant to overcome them.

Cash keeps you in the game. A working feedback loop helps you learn how to play it.

The distance between survival and understanding, between remaining able to choose and becoming slightly better at choosing, is where most of the real work of corporate finance happens. The company preserves enough room to continue, turns what happened into evidence, and tries again with a model that is a little less wrong than the one before it.