Chapter 02 · Part I · The case: Why decisions, not data or models
Pain and value
Whitepaper · Pain and value
1. Pain and value#
Where the pain in investment decisions comes from, what it costs, and why decision quality is the only durable alpha
Three roots, uncertainty, complexity and speed, against six fixed human limits; four families of inefficiency mapped to measurable business outcomes; decision quality as the measure that separates skill from luck; and analysis rethought beyond human limits.
| 01 · the pain | 02 · the value | 03 · rethinking |
|---|---|---|
| Beyond human limits | Priced, not asserted | Not the same analysis faster |
| Three roots, six human limits, four families of inefficiency. | Every inefficiency mapped to a business outcome and a value equation. | Every name, thousands of scenarios, every regime; the human decides. |
Summary
2. In brief#
The pain has three roots: uncertainty, complexity and speed. They are properties of markets, not of any firm. Each runs into a fixed human limit, and the best firms suffer them too, most quietly, because good outcomes hide them. Priced in the firm's own numbers, that pain is the value of decision optimization.
Three roots, and none of them is the firm's fault.
Uncertainty, complexity and speed are properties of markets. They run into six fixed human limits: working memory, attention, bandwidth, memory decay, statistical intuition and time. Experience changes what a person does inside the limit; it does not widen it.
The pain shows up as four families of inefficiency.
Governance, information, analysis and behavioural. Each family has named members with a cause, a symptom and a measure, from unwritten criteria and commodity information to backtest overfitting, fat tails, occasion noise and outcome bias.
Every inefficiency maps to a measurable business outcome.
Risk reduction is Δ probability × exposure × loss given event; effectiveness is decisions × capital × Δ decision quality; efficiency is hours × decisions × cost per hour. The value is underwritten in monetary value with the firm's own numbers.
A good decision is defined by its quality, not its outcome.
Decision quality is the quality of the end-to-end architecture that produces the outcome: frame, context, analysis, recommendation, process, record and learning, measured by calibration, policy adherence and conviction against outcome.
Decision quality separates skill from luck.
Signal alpha decays once it is shared. Decision alpha is specific to the firm, causal, repeatable and compounding, which makes decision quality the only durable source of alpha.
Beyond human limits, analysis is rethought, not automated.
Every name instead of a shortlist, thousands of scenarios instead of three, a calibrated distribution instead of a point, every regime at four scopes at once, and counterfactuals as a routine input.
Judgment becomes signal, and memory never forgets.
Expert judgment is captured as structured predictions, beliefs update continuously, and a decision memory keeps every decision made and not made. The human moves from producing analysis to judging, framing and owning the decision.
The argument behind the decision audit.
The four inefficiencies are the checklist the audit runs against a firm's decisions, the value equations are how the findings are priced, decision quality is the standard a decision is held to afterwards, and rethought analysis is what engineering aims for.
01
Where the pain comes from
3. Three roots, six limits, four inefficiencies#
Every active manager believes that their process, their people and their framework set them apart, and many are right about the people. The inefficiencies are the same in every firm because they do not come from the firm. They come from three properties of the world that run into six properties of the human mind.
The three roots#
Investment decisions are made under uncertainty: the outcome of any decision is one draw from a distribution that is fat tailed, skewed and dependent on a regime that can change. They are made in complexity: thousands of names, hundreds of drivers, interactions across sleeves and silos, each decision connected to the next one in the decision lifecycle. And they are made at speed: information arrives continuously, windows close in days and the number of decisions per person grows every year. None of these is new. What has changed is that all three have moved past the point at which a person, however able, can handle them by hand.
| Root | What IT is | The human limit IT exceeds | The inefficiency IT produces |
|---|---|---|---|
| Uncertainty | Outcomes are draws from distributions that are fat tailed, skewed and regime dependent. The same decision has many possible futures, and only one of them happens. | Statistical intuition: people think in points, stories and small samples, not in distributions. Memory decay: we remember what happened, not the probability we gave it. | Analysis (point estimates, Gaussian assumptions, small samples), behavioural (overconfidence, outcome and hindsight bias), governance (review by outcome). |
| Complexity | Thousands of names, hundreds of drivers, interactions across sleeves, silos and stakeholders, each decision connected to the next in the decision lifecycle. | Working memory holds about four chunks at once. Attention is serial. Bandwidth is hours per analyst. A rule too large to hold in one head stays unwritten. | Information (coverage, silos, unseen alternatives, hidden variables), analysis (low dimensionality, no counterfactuals), governance (unwritten criteria, drift). |
| Speed | Information arrives continuously, windows close in days, cycles shorten, and the number of decisions per person rises every year. | Time: an update by hand takes hours; data arrives in seconds. Fatigue and pressure: judgment degrades through the day and under a deadline. | Behavioural (occasion noise, time pressure, FOMO), information (stale data, calendar refresh), governance (drift across cycles, undocumented shortcuts). |
The six human limits#
The limits are well documented and they do not move with talent, experience or effort. Experience changes what a person does inside the limit; it does not widen it. An experienced investor has better chunks, not more of them.
Why this is not a verdict on the firm#
A firm with excellent people still has four chunks of working memory per person, still updates its views when the calendar says, still remembers outcomes better than probabilities and still cannot read 3,000 names by Friday. The inefficiencies of chapter 02 are the predictable result of asking people to do what people cannot do, and the firms that suffer them least are not the ones with the best people but the ones that have moved the work past the limit into a system, while keeping the judgment with the people.
a human limit, not a firm failure. Every symptom in chapter 02 passes this test.
02
The four inefficiencies · Drill down
4. The four inefficiencies#
Four families, one per source of alpha. Each member is named, traced to its root and its limit, shown as it appears in a firm, and given the symptom by which the decision audit measures it. The families are the same at every level: solution, decision policy and analysis.
Governance inefficiencies are the gap between the strategy the firm believes it runs and the one it actually runs. They come from complexity (the full rule is too large to hold in a head), from speed (writing things down costs hours nobody has) and from uncertainty (the outcome is remembered, the probability is not). Their common symptom is that the firm cannot show, to itself or to an investor, that its edge is applied the same way every time.
| Inefficiency | What IT is | Where IT comes from | How IT shows up | Measurable symptom |
|---|---|---|---|---|
| Unwritten criteria | The mandate's real criteria, weights and thresholds exist in heads and habits, not in a policy that can be applied, tested or taught. | Complexity, working memory: a full rule is too large to hold, so it is simplified differently by each person. | Two analysts rank the same universe differently and both believe they follow the strategy. New hires learn by imitation. | Share of key decisions with a written policy; rank agreement between analysts on the same cases. |
| Drift between analysts, funds and cycles | The same criteria are applied with different weights by different people, in different funds and at different points of the cycle. | Speed and complexity: no one sees all the cases at once, and the cycle changes the mood in which they are judged. | Style drift that only the allocator notices. Risk taken in year three that would have been refused in year one. | Dispersion of scores on identical cases; drift of effective weights over time; style drift against the stated strategy. |
| Key-person dependency | The decision logic lives in one or two people. When they leave, are ill or are busy, the edge leaves with them. | Complexity: the logic was never externalised because it was too large and too tacit to write down. | The firm cannot scale a strategy to a second team. Allocators price the key-person risk into the allocation. | Number of decisions that cannot proceed without one named person; share of logic documented in a policy. |
| Outcome-based review | Decisions are judged by their result, not by the quality of the decision at the time it was made. | Uncertainty: the outcome is the only thing that is visible and remembered; the probability was never recorded. | The lucky are promoted, the unlucky are blamed, and a flawed thesis that paid off is repeated with more capital. | Share of reviews that compare the outcome with the recorded probability; share of decisions with a recorded probability at all. |
| No audit trail | The data, drivers, model version, options and reasons behind a decision are not kept in one place at the time of the decision. | Speed: writing it down costs hours nobody has; memory decay: it is reconstructed later, if at all. | Due diligence questionnaires answered from memory. Regulators and investors are shown the deck, not the decision. | Share of decisions with a complete record at decision time; hours per due diligence request. |
| Decisions not reconstructable | Nobody can say, a year later, why a decision was made, what else was considered and what would have changed it. | Memory decay and hindsight: what is recalled is a story consistent with what happened. | Post-mortems that explain the result rather than test the decision. Lessons that are the wrong lessons. | Share of past decisions that can be replayed with their context, options and rationale. |
| Inconsistent application of the mandate | Limits, exclusions and constraints are checked after the fact, or differently by each team, or waived under pressure. | Speed and attention: the constraint that is not in front of the person at the moment of decision is not applied. | Breaches found by risk or compliance. Overrides without a recorded reason. The mandate the investor bought is not the one that is run. | Breaches per year; overrides without reason; policy adherence (share of decisions made as the policy says). |
Information inefficiencies are the gap between what the firm could know and what it has in view when it decides. The central one is commodity information: the terminals, filings, broker notes and vendor datasets that everyone has are priced before the memo is written, so research built on them reads like the consensus and positions built on them crowd with the peers who read the same screens. The firm's own context, which is where an edge could be, is expensive to assemble by hand, so bandwidth, silos and the calendar set what is seen.
| Inefficiency | What IT is | Where IT comes from | How IT shows up | Measurable symptom |
|---|---|---|---|---|
| Commodity information | Everyone has the same terminals, filings, broker research and vendor datasets. Information that everyone has is priced and carries no edge. | Complexity and speed: the firm's own, specific context is expensive to assemble, so the shared feed stands in for it. | Research that reads like the consensus. Positions crowded with the same peers who read the same screens. | Overlap of holdings and theses with peers; share of inputs unique to the firm; crowding measures. |
| Coverage limited by bandwidth | The universe is as wide as headcount times hours. What is not covered is not seen, and what is not seen is neither bought nor avoided. | Bandwidth: a person reads at a fixed rate; the universe grows regardless. | The shortlist is the familiar list. The best opportunity of the year was in a name nobody had time to open. | Names or deals assessed per analyst per year; share of the universe scored on every criterion of the mandate. |
| Silos | Data, views and risks sit in separate teams, sleeves, funds and systems. The driver that connects them is in none of them. | Complexity and attention: each team can only attend to its own part; the whole is nobody's job. | The same macro exposure is held three times under three labels. A credit team learns of an equity team's thesis break from the price. | Hidden concentration (positions sharing one driver); time for a signal to travel between teams. |
| Stale and point-in-time errors | Views are built on data as it is today, not as it was known at the time; refreshes follow the calendar, not the evidence. | Speed: data arrives continuously, updating by hand does not. | Backtests that know the restated figure. Reviews that run on last quarter's numbers. Decisions taken on a view that was already wrong. | Share of analyses run on point-in-time data; age of the data behind a decision at decision time. |
| Look-ahead and survivorship bias in data | Datasets that only contain the survivors, and histories that quietly include information that was not available at the time. | Statistical intuition: the bias is invisible in the result and nobody has time to check the vintage of every field. | Strategies that work in the backtest and not in the book. Base rates built on the companies that did not fail. | Vintage checks per dataset; survivorship-adjusted base rates; backtest against live gap. |
| Unseen alternatives | The options that were never generated: the names not screened, the structures not considered, the actions not priced. | Working memory and bandwidth: options are generated one at a time and the list stops when time runs out. | Two options in the memo, one of them the straw man. The counterfactual of not acting is never costed. | Options per decision; share of decisions with the do-nothing and the deferral priced. |
| Hidden variables | The driver that explains the outcome is not in the model, the screen or the sector label, because it sits outside the classification. | Complexity: real drivers cross sectors, regions and asset classes; classifications are made to fit in a head. | A portfolio diversified by sector that moves as one on rates or on one customer. Attribution that cannot explain half of the return. | Share of return attributed to identified drivers; unexplained variance; concentration on latent drivers. |
| Noise in the signal | Most of what moves on a screen is noise; the few changes that move the odds of a thesis are buried in it. | Attention and speed: a person alerted on everything attends to nothing. | Hundreds of alerts, none acted on. A thesis break noticed when the price already shows it. | Precision and recall of alerts; lead time of a trigger before the event. |
Analysis inefficiencies are the statistical limitations of doing inference by hand on a few points of a non-stationary, fat-tailed process. They are not errors of carelessness; most of them are made by careful people using the tools that fit in a spreadsheet and a head. The list below is the one the audit checks every analysis against.
| Inefficiency | What IT is | Where IT comes from | How IT shows up | Measurable symptom |
|---|---|---|---|---|
| Small samples and non-stationarity | Markets provide few independent observations, and the data-generating process changes under them. A decade of monthly data is 120 points from several different worlds. | Uncertainty: the world is non-stationary; statistical intuition treats a short history as a law. | Relationships that held for ten years break in a week (bonds and equities falling together in 2022). Rules fitted on one regime run in another. | Out-of-sample against in-sample accuracy; stability of estimated relationships across regimes. |
| Overfitting and backtest overfitting | A model or strategy fitted to the one history that happened learns its noise. The more variants are tried, the better the best one looks by chance. | Speed and bandwidth: trying many variants is cheap; testing each one honestly is slow. | The backtest Sharpe ratio is twice the live one. The best-looking strategy of the year was the luckiest, not the best. | Live against backtest performance; number of variants tried per strategy adopted; deflated Sharpe ratio. |
| P-hacking of factors | Hundreds of published factors, most of which would not survive a multiple-testing correction; many decline sharply after publication (Harvey, Liu and Zhu 2016; McLean and Pontiff 2016). | Uncertainty and commodity information: the same data mined by thousands of researchers yields patterns that are chance or already arbitraged. | Factor tilts bought at the top of their popularity. Signals that stop working the year they are adopted. | Post-adoption decay of signals; share of signals with a causal rationale rather than a statistical one. |
| Linear and Gaussian assumptions against fat tails and skew | Models assume normal, symmetric, linear relationships. Returns are fat tailed and skewed; most single stocks lose to treasury bills and a few make all the return (Bessembinder 2018). | Statistical intuition: the bell curve is the only distribution most people can picture; linear is the only relationship a spreadsheet draws easily. | Value at risk that is breached far more often than its confidence level says. Portfolios built for the average that die in the tail. | Tail coverage of forecasts (share of realised outcomes outside the 5 to 95 per cent interval); skew of realised against assumed distributions. |
| Correlation mistaken for causation | Co-movement is read as a driver. Returns are attributed to the wrong cause, so the wrong lever is pulled next time. | Complexity: causal structure takes many variables and explicit assumptions; correlation takes one line of code. | Leverage-driven private equity returns credited to operational skill. A hedge that correlated in calm markets and not in the crash. | Share of forecasts with a stated causal driver; attribution tested against the counterfactual, not the plan. |
| Point estimates instead of distributions | The base case is the forecast. The upside and downside, if present, are two more points, not a distribution with its probabilities. | Statistical intuition and working memory: one number can be held, discussed and remembered; a distribution cannot. | Committees that decide on the base case and are surprised by the downside. Conviction that cannot be checked because it was never a probability. | Share of forecasts issued as calibrated distributions; interval coverage; Brier score. |
| Low dimensionality | Models carry the few drivers a person can keep in view. Everything else is in the error term, where it becomes the surprise. | Working memory and bandwidth: every added driver costs time to source, model and explain. | A credit model with six ratios on an issuer whose fate is decided by one customer and one regulation. | Drivers per analysis; unexplained variance; share of outcomes attributed to drivers outside the model. |
| No counterfactuals Only the history that | happened is analysed. What would have happened under another action, or in the worlds that did not occur, is not computed. | Uncertainty: counterfactuals are not observable; computing them needs a causal model and simulation, not a spreadsheet. | A credit book judged safe because it survived a mild slowdown. A manager judged on one path of many. | Share of decisions evaluated against a counterfactual; regret measured against the best action ex post. |
| Calendar-driven updating | Views are refreshed at the review date, not when the evidence changes. Between reviews the decision runs on a stale posterior. | Time: updating by hand costs hours, so it is rationed by the calendar. | Quarterly reviews. A thesis that broke in week two is carried for eleven more. | Time from evidence to updated view; share of updates triggered by evidence rather than date. |
| Regime blindness | Analyses assume the regime that is current at the four scopes: global, local, macro (the economic cycle) and micro (the investable universe, the corporate lifecycle). The change of regime is seen afterwards. | Uncertainty and complexity: regimes are latent, overlap across scopes and are identified by many weak signals at once. | Growth rules run into a value regime. A company treated as mature that is in its decline. Cycle-top leverage. | Regime identified before the turn (lead time); outcome in the worst regime against the average. |
Two numbers that put the statistics in perspective. With a true annual information ratio of 0.5, a track record needs about 16 years before its t-statistic reaches 2 (t = IR × √T). And of the hundreds of factors published in the academic literature, most would not clear the multiple-testing hurdle that Harvey, Liu and Zhu (2016) propose, while McLean and Pontiff (2016) find that signal returns fall substantially after publication. Analysis built on small samples and shared signals is not an edge; it is the consensus with a confidence interval nobody computed.
Behavioural inefficiencies are what the limits do to judgment itself. Following Kahneman, Sibony and Sunstein (2021), the first group is noise: unwanted variability in judgments that should be identical. It has three components, level, pattern and occasion noise, and in every professional setting that has been audited it is larger than the people in it expect. The second group is bias: systematic error in one direction. The third is what overload, time pressure and the absence of a calibration feedback loop do to conviction, sizing and the path of decisions over time.
Noise, overload and conviction
| Inefficiency | What IT is | Where IT comes from | How IT shows up | Measurable symptom |
|---|---|---|---|---|
| Level noise | Different people judge the same case at systematically different levels: one underwriter is strict, another lenient, on every case. | Complexity: there is no shared scale, so each person brings their own. | One deal partner's cases always pass; another's never do. Scores depend on who scored. | Dispersion of scores between judges on the same cases. |
| Pattern noise | The same person weighs criteria differently from case to case, by idiosyncratic reactions to features of the case. | Working memory and unwritten criteria: with no fixed weights, each case triggers a different subset. | An analyst strict on leverage in one case and relaxed in a similar one, with a story for each. | Within-judge inconsistency on matched cases; noise audit (Kahneman, Sibony and Sunstein 2021). |
| Occasion noise | The same person judges the same case differently on another day, depending on mood, fatigue, time of day and what came before. | Speed and fatigue: judgment degrades through the day and under pressure. | The committee approves on Monday morning what it would reject on Friday afternoon. The deal after a loss is judged harsher. | Test-retest consistency; order and time-of-day effects in committee outcomes. |
| Overload and time pressure in committees | Too many cases, too much material, too little time. The committee attends to the narrative and the presenter rather than the evidence. | Attention and time: serial attention under a clock. | Forty-page memos read in the meeting. The last item on the agenda is decided in five minutes. | Cases per committee hour; share of committee time on questions that change the decision. |
| Miscalibrated conviction | Confidence is expressed in words (high conviction) that were never tested against outcomes, so nobody knows whether high means 60 or 90 per cent. | Uncertainty and memory: the probability is never written down, so it is never checked. | Everyone is confident. Nobody is calibrated. Position sizes follow the loudest voice. | Calibration curve per person (Brier score, reliability by bucket). |
| Sizing by rule of thumb | Positions are sized by habit, equal weight, or the strength of the feeling, not by conviction, codependency and contribution to the book. | Complexity: joint distributions cannot be held in a head; a rule of thumb can. | Selection alpha lost in sizing. Hidden concentration on one driver under ten labels. | Sizing contribution against equal weight; realised against intended risk. |
| Path dependency | Early, unexamined choices lock in later options; positions are kept because they were taken, and errors compound along the path. | Memory and loss aversion: what was decided becomes the anchor for what is decided next. | A thesis defended long after its driver has gone. An add-on bought to justify the platform. | Share of positions re-underwritten from zero; time to exit after a recorded thesis break. |
Biases, and where each one bites
| Bias | What IT is | Where IT bites in investing | Measurable symptom |
|---|---|---|---|
| Overconfidence | Intervals too narrow, probabilities too extreme, own skill overrated. | Forecast ranges that are breached half the time; leverage sized on the base case. | Interval coverage far below the stated level. |
| Anchoring | The first number (the ask, the last comparable, the screen price) pulls the estimate. | Prices set from the seller's ask; targets that stay near the IPO price. | Dispersion of valuations with the anchor removed. |
| Availability | What comes easily to mind (recent, vivid, reported) is judged more likely. | Risk discussions dominated by last year's crisis; the rare scenario dropped. | Scenario sets that track the headlines of the quarter. |
| Confirmation | Evidence for the thesis is sought and weighted; evidence against it is explained away. | Memos that cite ten supporting and no opposing facts; thesis breaks denied. | Ratio of disconfirming evidence in memos; time to act on a break. |
| Loss aversion | Losses weigh roughly twice as much as equal gains. | Hedges bought after the loss; winners cut, losers held. | Asymmetry of holding periods for gains and losses. |
| Disposition effect | Selling winners too early and holding losers too long. | Realised returns below the paper returns of the same ideas. | Realised against hypothetical hold returns. |
| Herding | Following peers and consensus because being wrong alone is worse than being wrong together. | Crowded trades; the benchmark hugged; the same add-on bid by five sponsors. | Overlap with peers; active share; crowding score. |
| FOMO | Fear of missing out drives entry at the point of maximum attention and price. | Late-cycle deals at peak multiples; thematic entries after the move. | Entry timing against the attention peak; multiple paid against cycle average. |
| Sunk cost | Past investment of money, time or reputation justifies further investment. | Follow-on rounds into failing companies; the deal that cost too much to drop. | Follow-on decisions scored as new decisions against the counterfactual. |
| Narrative fallacy | A coherent story feels true; the data that does not fit the story is discarded. | Equity stories that survive their numbers; the macro narrative that explains everything after the fact. | Share of thesis claims with a tested causal driver. |
| Hindsight bias | After the outcome, it seems to have been predictable; the probability at the time is forgotten. | Post-mortems that find the obvious warning sign; analysts blamed for what nobody saw. | Recorded probability at decision time against the recalled one. |
| Outcome bias | The decision is judged by its result rather than by what was known at the time (Baron and Hershey 1988). | The reckless success promoted; the sound loss punished; a flawed thesis repeated with more capital. | Share of reviews that score decision quality separately from outcome. |
Limits, not failures#
None of the above is a finding against the people. Noise is largest in the most experienced committees because experience adds conviction without adding calibration. Hidden variables are most dangerous in the best-diversified books because diversification by label hides concentration by driver. Outcome-based review hurts most in the firms with the best track records, because a good record provides the most cover for the lucky decisions in it. The best firms suffer the four inefficiencies most quietly: their outcomes hide the symptoms until the regime changes, and then the post-mortem, run by outcome and from memory, finds the wrong lesson.
What follows from this. If the inefficiencies come from limits, they are not removed by hiring better people, by a stricter process or by more effort. They are removed by moving the work that exceeds the limit into a system that does not have it, and by keeping with the people the judgment, the framing and the accountability that a system cannot carry. Chapter 05 describes what that looks like. Chapter 03 first says what it is worth.
03
From inefficiency to measurable business outcomes
5. What the pain costs, and how to measure it#
An inefficiency is a cost only once it is attached to a business outcome and an equation. The equations below are the ones the decision audit fills with the firm's own numbers (Template V); the worked examples show the order of magnitude for three typical firms. Every number in them is illustrative.
Seven benefits, seven equations#
Benefits are the forms value takes: efficiency, speed, scale, risk reduction, decision quality, effectiveness, innovation, transparency, learning and fundraising success, the ten measures of the buyer's guide under the six forms of benefit of the delivery framework (efficiency, scale, coverage, effectiveness, experience, innovation). Seven of them can be written as an equation with inputs the audit collects; decision quality is their common cause and is measured in chapter 04; innovation is what could not be done before and is valued per use case.
| benefit | equation | inputs the audit collects |
|---|---|---|
| Risk reduction | Δ probability of the loss event × exposure × loss given event | Loss events of the past cycle; exposure per event; share of events that a wider scenario set or an earlier trigger would have caught. |
| Effectiveness | Decisions per year × capital per decision × Δ decision quality (in return) | Decision count by type; average capital moved; dispersion between analysts and against the counterfactual decision on the same cases. |
| Efficiency | Hours freed per decision × decisions per year × cost per hour | Hours per screen, memo, review and due diligence request; loaded cost per hour. |
| Speed | Opportunities priced before action × value lost per late action | Time from signal to decision; price move during that time; windows missed. |
| Scale | Coverage per person after ÷ before × value of coverage (hit rate × value per hit) | Names or deals assessed per analyst; shortlist hit rate; value per successful pursuit. |
| Transparency and fundraising | Δ probability of a successful raise or mandate × size × fee margin, plus due diligence hours saved | Due diligence outcomes and questions; time per questionnaire; allocator feedback on process evidence. |
| Learning | Δ calibration per cycle × decisions per cycle × capital per decision | Calibration curves per analyst and per analysis, by cycle; policy changes made on evidence. |
How a risk becomes a loss#
The risk equation deserves one paragraph because it is where most of the money is. A missed tail scenario is not a loss until the scenario happens; it is an expected loss, the probability of the event times the exposure times the loss given the event. Decision optimization changes the first term by seeing the scenario before capital is committed (the credit book that survived a mild slowdown and was never tested against a liquidity freeze), and the third term by acting earlier when the event starts (a trigger on drivers rather than on price, ten days and three per cent before the review would have found it). Both are measurable after the fact: the share of realised loss events that were in the scenario set, and the lead time of the trigger before the event.
Inefficiency, outcome, loss or gain, measure, KPI#
The table maps every member of the four families to the outcome it affects, an example of the loss it causes or the gain its removal brings, how the audit measures it and the KPI that proves it afterwards.
| Inefficiency | Outcome affected | Example loss or gain | How to measure | KPI |
|---|---|---|---|---|
| Governance · unwritten criteria, drift | Effectiveness, decision quality | Style drift costs the mandate: the allocator leaves, or risk taken in year three that the policy would have refused. | Dispersion of scores between analysts on the same cases; drift of effective weights over time. | Policy adherence; rank agreement |
| Governance · key-person dependency | Scale, fundraising | A second team cannot run the strategy; the raise is discounted for key-person risk. | Decisions that need one named person; share of logic in policies. | Decisions runnable without the key person |
| Governance · outcome-based review, no record | Learning, decision quality | A lucky flawed thesis is repeated with twice the capital; the sound unlucky analyst leaves. | Share of reviews scored on decision quality; share of decisions with a recorded probability. | Calibration; records complete at decision time |
| Governance · inconsistent mandate | Risk reduction, transparency | Breaches found by compliance; a mandate lost after an audit; overrides that cannot be explained. | Breaches and overrides per year, with and without reason. | Breaches; overrides with reason |
| Information · commodity information | Effectiveness | Crowded positions unwind together; the edge that everyone has is paid for and not earned. | Overlap with peers; share of inputs unique to the firm. | Crowding; unique inputs per thesis |
| Information · coverage, unseen alternatives | Scale, effectiveness | The best deal of the vintage was in a name nobody opened; the do-nothing option was never priced. | Universe scored per analyst; options per decision; shortlist hit rate. | Coverage per person; hit rate |
| Information · silos, hidden variables | Risk reduction | The same driver held three times under three labels; a 15 per cent drawdown on one customer or one rate path. | Positions sharing one driver; unexplained variance in attribution. | Hidden concentration; attributed share of return |
| Information · stale data, look-ahead, survivorship | Risk reduction, effectiveness | A strategy that worked in the backtest and not in the book; a decision taken on last quarter's numbers. | Backtest against live gap; age of data at decision time; vintage checks. | Point-in-time share; live against backtest |
| Information · noise in the signal | Speed, risk reduction | A thesis break noticed on the price, 10 days and 3 per cent late. | Precision and recall of alerts; lead time of triggers. | Lead time; alert precision |
| Analysis · small samples, overfitting, p-hacking | Effectiveness, risk reduction | The signal that stopped working the year it was adopted; a Sharpe ratio that halves live. | Out-of-sample accuracy; signal decay after adoption; variants tried per strategy. | Live against backtest; signal half-life |
| Analysis · Gaussian, point estimates, low dimensionality | Risk reduction | Value at risk breached ten times more often than it should; the downside nobody put a number on. | Interval coverage; Brier score; drivers per analysis. | Tail coverage; calibration |
| Analysis · correlation, no counterfactuals, regime blindness | Effectiveness, learning | The wrong lever pulled next time; a hedge that failed in the crash; cycle-top leverage. | Attribution against the counterfactual; outcome in the worst regime; regime lead time. | Attributed cause; worst-regime outcome |
| Analysis · calendar-driven updating | Speed | A broken thesis carried for eleven weeks; a window closed before the review. | Time from evidence to updated view. | Evidence-to-update time |
| Behavioural · noise (level, pattern, occasion) | Decision quality Monday's approval is Friday's | rejection; the strict underwriter and the lenient one on the same book. | Noise audit on matched cases; test-retest consistency. | Dispersion between and within judges |
| Behavioural · biases | Effectiveness, risk reduction | Late-cycle entries at peak multiples; losers held, winners sold; the ask as the anchor. | Entry timing against attention; holding-period asymmetry; valuation dispersion without anchor. | Timing regret; disposition measure |
| Behavioural · overload, miscalibrated conviction, sizing | Decision quality, effectiveness | Selection alpha leaked in sizing; conviction that means 60 per cent in one mouth and 90 in another. | Calibration curve per person; sizing contribution against equal weight. | Brier score by person; sizing skill |
| Behavioural · path dependency | Risk reduction, learning | The add-on bought to justify the platform; the thesis defended a year after its driver left. | Share of positions re-underwritten from zero; time to exit after a recorded break. | Time to exit after break |
Reading the table. The same inefficiency can surface in several outcomes, and the same outcome can have several causes; the audit attributes by the four families so that the value equation per use case counts each only once. Benefits arrive at different speeds: efficiency, speed and transparency within weeks; decision quality at the analysis level within months; effectiveness at fund level over years. The evaluation plan uses the fast evidence early and lets the slow evidence build.
Three worked examples#
The profiles are types, not clients. The deltas (4 percentage points, 0.3 turns, 25 bps, 10 days) are the kind of number the audit establishes from the firm's own past decisions and the proof of value then verifies; they are placeholders here. Effectiveness is always measured against the counterfactual decision the old process made or would have made on the same cases.
04
What makes a good decision
6. Decision quality, skill and luck, durable alpha#
A good decision is not a decision that turned out well. It is a decision made the right way, judged against the probabilities it was made under, and the only kind that can be repeated, improved and shown to an investor as skill.
DEFINITION
Decision quality
The quality of the end-to-end architecture that produced an outcome: the frame and objective, the context, the analysis, the recommendation, the process, the record and the learning, judged against the probabilities the decision was made under rather than by the outcome alone. It is measured by calibration, policy adherence and conviction against outcome.
Seven dimensions, three measures#
| Dimension | What good looks like | How IT is measured | Alpha |
|---|---|---|---|
| Frame and objective | The decision is the right decision, with one clear objective, the mandate's goals and constraints, and its place in the decision lifecycle explicit. | Objective and criteria written before the analysis; decision type (Selection, Policy optimization, Go / No-Go, Monitoring) assigned. | Governance |
| Context | Every relevant name, driver, regime and comparable situation in view, point in time, beyond what the team had time to read. | Coverage; point-in-time share; drivers per analysis; age of data at decision time. | Information |
| Analysis | Causal diagnosis and counterfactual prediction that yield calibrated distributions, not points; every regime at four scopes; the tails explicit. | Calibration (Brier score, reliability by bucket); interval coverage; out-of-sample accuracy. | Analysis |
| Recommendation | Options, conviction and rationale explained, ranked against the mandate, within its limits, with what would change the answer. | Share of recommendations with options, conviction and explanation; limit breaches at recommendation time. | Analysis, Governance |
| Process | The policy followed; conviction recorded as a probability; the decision owner decides; overrides recorded with a reason. | Policy adherence; overrides with reason; cases per committee hour on the questions that matter. | Governance, Behavioural |
| Record | The context, options, prediction, conviction, decision and reason kept at decision time, so the decision can be replayed. | Share of decisions with a complete record at decision time. | Governance |
| Learning | Outcomes compared with the recorded prediction; variance attributed by cause along the causal decision graph; calibration fed back to people and policies. | Calibration improvement per cycle; policy changes made on evidence; conviction against outcome. | Behavioural, Analysis |
CALIBRATION
Whether stated probabilities hold across many decisions: of the decisions labelled about 70 per cent, did about seven in ten succeed? Measured per analysis, per decision type and per contributor, by Brier score and reliability by bucket. Checkable within months, long before the fund result is in.
POLICY ADHERENCE
The share of decisions made as the policy says, with every override recorded with its reason. It is the process measure: a policy that is not followed cannot be evaluated, and an override without a reason cannot be learned from.
CONVICTION AGAINST OUTCOME
Whether high-conviction decisions did better than low-conviction ones, per person and per policy. It is what turns a word (high conviction) into a number that can size a position and can be checked.
Skill and luck#
Outcomes are a noisy measure of decisions. A sound call can lose and a careless one can win, and in investing the noise is large: the dispersion of outcomes around the expected value of a single decision is many times the edge in the decision. Judging a decision by its result is outcome bias. Baron and Hershey (1988) showed that people rate the very same decision, made on identical information, as worse when its result was bad. Firms that judge by results therefore reward the lucky, punish the unlucky and, most expensively, reinforce the flawed thesis that happened to pay off.
Three facts make the separation of skill from luck a matter of system rather than of judgment. First, a single outcome cannot show that a probability was wrong; a loan approved at 72 per cent that fails is one of the possible worlds, with a 28 per cent weight. Second, the number of decisions needed to tell skill from luck by outcome alone is large: with a true information ratio of 0.5, about 16 years of track record for a t-statistic of 2, and most strategies do not run unchanged that long. Third, people cannot keep the probability they gave at the time once they know the outcome; hindsight rewrites it. So the probability has to be recorded at decision time, the calibration has to be computed across many decisions, and the outcome has to be attributed by cause along the causal decision graph: market, factor, thesis, timing, execution and the residual that is luck. That is what the Learning capability does, and it is why decision quality can be shown within months at the analysis level while the fund outcome takes years.
Why decision quality is the only durable source of alpha#
SIGNAL ALPHA DECAYS
An information or analysis edge that others can find is found. McLean and Pontiff (2016) show that the returns of published predictors fall substantially after publication; Harvey, Liu and Zhu (2016) show that most published factors would not survive a multiple-testing correction in the first place. Commodity data, shared models and the same screens produce the same positions, and the edge is competed away or was never there.
DECISION ALPHA COMPOUNDS
A better way of deciding is specific to the firm: its mandate, its context, its experts, its decision memory. It is causal, because it rests on an explicit model of what moves what; repeatable, because it is a policy and not a person; and compounding, because every decision improves the context, the models and the calibration for the next. It cannot be arbitraged away by a competitor who does not have the firm's decisions.
The four alphas are the four sources of decision alpha, and each one improves particular dimensions of decision quality. The mapping is what lets the audit attribute a decision quality gap to its source and the proof of value show which source delivered.
| Alpha | What IT is | Decision quality dimensions | Measured by |
|---|---|---|---|
| Governance alpha | The edge applied consistently across analysts, funds and cycles, with a record. | Frame and objective, process, record | Policy adherence; records complete; dispersion between judges |
| Information alpha | Context beyond manual bandwidth and consensus; hidden variables surfaced earlier. | Context | Coverage per person; unique inputs; lead time |
| Analysis alpha | Causal, probabilistic, counterfactual and regime-aware analysis instead of extrapolation. | Analysis, recommendation | Calibration; interval coverage; attributed cause |
| Behavioural alpha | Less noise, less bias, less overload; conviction calibrated and sizing tied to it; every outcome learned from. | Process, learning | Brier score by person; sizing skill; calibration improvement per cycle |
Proof of skill. An allocator who sees a track record sees one path through a noisy process. An allocator who sees the decision record, with every decision's probability at the time, its calibration across hundreds of decisions, its policy adherence and the attribution of outcomes by cause, sees skill separated from luck. That evidence exists in months, it survives the departure of any one person, and it is the answer to the due diligence question that a track record cannot answer: would it happen again?
05
Beyond human limits · Rethinking analysis
7. Not the same analysis faster: a different analysis#
Most AI projects in investing take today's process and make it faster: summarize the memo, extract the data, draft the report. If the process runs on correlations, three scenarios, point estimates and outcome-based review, automation produces the same errors at higher speed. The gain lies elsewhere: in what becomes possible once the six limits no longer set the shape of the analysis.
The argument#
Today's analysis has the shape of the limits that produced it. A shortlist exists because one person can only open thirty names. Three scenarios exist because a committee can hold three. A point estimate exists because one number fits in a sentence. A quarterly review exists because updating by hand takes a quarter. Sector labels exist because real drivers do not fit in a head. The memo has two options because generating a third took the afternoon. None of this is the analysis an investor would design if the limits were not there; it is the analysis that fits through them.
Automating this analysis keeps its shape. The shortlist is produced faster, the three scenarios are drafted by a model, the point estimate arrives in the morning. The limit has moved from the person's bandwidth to the person's habits, and every inefficiency of chapter 02 survives. Rethinking the analysis starts from the other end: from the decision, the question it answers and the mandate it serves, and asks what analysis would best answer that question if coverage, dimensionality, scenario count, updating frequency and memory were no longer constrained. The answer is not a faster version of the old analysis. It is a larger and different one, and it changes what the people in the firm do.
| What becomes possible | Today, within human limits | Rethought, beyond human limits |
|---|---|---|
| Every name, not a shortlist | An analyst opens 30 names in depth; the universe of 3,000 is screened by label and habit. | Every name scored on every criterion of the mandate, from its drivers, with a probability of being the best; the human reads the top of a ranking that explains itself. |
| Thousands of scenarios, not three | Base, upside, downside: three points chosen by the author of the memo. | Thousands of simulated paths from a causal model with the tails, the correlations and the regime changes inside them; the question becomes under which paths the case breaks. |
| A calibrated distribution, not a point | One number, a target price, a base-case IRR, a default view. | A distribution with its 5 to 95 per cent interval and a conviction label whose meaning has been checked against outcomes. |
| Every regime at four scopes at once | The current regime assumed; the change seen afterwards. | Global, local, macro and micro regimes identified continuously from many weak signals; the analysis carries a regime probability, and the recommendation is tested across them. |
| Counterfactuals as a routine input | What happened is analysed; what would have happened is a conversation. | What would happen under each action, and what would have happened under the one not taken, computed before capital is committed and after the outcome is in. |
| Expert judgment as structured predictions | Views in prose; conviction in adjectives; the loudest voice wins. | Experts answer precise questions with a range and a confidence; their inputs are weighted by earned calibration and fused with models by regime. |
| Causal diagnosis, not correlation | A regression on the drivers that fit in a spreadsheet. | An explicit causal decision graph from experts and data, tested against outcomes by root cause; actions can be simulated because the graph says what moves what. |
| Continuous in-context updating | The quarterly review; the thesis carried between reviews. | Every holding watched on its drivers; the posterior updated as evidence arrives; a trigger fires only when the odds of the thesis change. |
| A decision memory that never forgets | Memos in a drive; reasons in heads; the decisions not taken nowhere. | Every decision with its context, options, prediction, conviction and reason; the decisions not taken scored against what happened, so the firm learns from both. |
| The whole decision lifecycle, not one decision | Each decision optimized alone; value lost at the hand-off to the next. | Selection, Go / No-Go, price, sizing, timing, financing, monitoring, hedging, exit connected, so each decision carries what the last one learned. |
Two of these deserve a sentence more. Expert judgment is not replaced: it is captured as structured predictions, a range and a confidence on a precise question, weighted by the expert's earned calibration and fused with models by regime, so that it counts most exactly where the data is weakest and the expert has been right before. The decisions not taken are the half of the firm's experience that no memo records; a decision memory that scores them against what happened is the only way to learn whether the firm's no was right as often as its yes.
Before and after, per analysis component#
An analysis has four components, in the order of the intelligence flow: Perceive context, Causal diagnosis, Counterfactual prediction and Contextual optimization. Each is rethought, not accelerated.
| Analysis component | Today, within human limits | Rethought, beyond human limits |
|---|---|---|
| Perceive context | A data room assembled by hand for the names on the shortlist; sector labels; data as of today. | Context constructed on an ontology for every name, point in time, across sources and silos; comparable situations found by driver, not by label. |
| Causal diagnosis | Correlations, ratios and the author's story about what drives the asset. | An explicit causal decision graph of drivers, fitted from data and experts, tested by root cause against outcomes, reusable across decisions as a decision asset. |
| Counterfactual prediction | A base case and two alternatives, Gaussian where a distribution is used at all, in the current regime. | Thousands of paths with fat tails, skew and regime change; what will, could and would happen; experts and models fused by regime; calibrated and checked. |
| Contextual optimization | The option in the memo, sized by rule of thumb, checked against limits afterwards. | The full action and term space searched against the distribution and the mandate's own goals and constraints, with the ripple-through to the rest of the book, limits applied before the recommendation. |
Before and after, per decision policy type#
Every decision policy has one of four types. The rethought analysis changes what each type produces for the decision maker.
| Decision policy type | Today, within human limits | Rethought, beyond human limits |
|---|---|---|
| Selection (selection, ranking) | A shortlist from a screen and from what the team knows; criteria weighted differently by each analyst. | The whole universe ranked on every criterion with a probability of being the best; the ranking records why each name sits where it sits and what was dropped. |
| Policy optimization (price, sizing, timing, allocation, financing, hedging, treatment) | One structure, one price, one size: the one that was proposed, anchored on the last comparable. | The optimum against the distribution and the walk-away, with the cost of waiting and the stress paths explicit, inside the policy's limits. |
| Go / No-Go (conviction test) | A yes or no on a base case and a narrative; the probability never stated; the threshold never written. | A calibrated probability against an explicit threshold, with the scenarios under which the case breaks and the terms under which it holds; every pass and override recorded. |
| Monitoring (trigger design, root cause) | Calendar reviews; alerts on price; post-mortems on outcome. | Triggers on the probability of the event, with lead time; root cause along the causal decision graph that separates market, thesis, timing, execution and luck. |
How it changes the firm#
This is augmentation, with the human deciding. The human / AI mode is set per component of each analysis: Automated, Augmented or Judgement only. What moves into the system is the work that exceeded the limits: scanning, remembering, simulating, keeping track, updating. What stays with the people is what the system
cannot carry: purpose, framing, judgment in new situations, the override and the accountability. The result is not fewer people; it is people doing a different job, and a firm that can show how it decides.
| What changes | Today | Rethought |
|---|---|---|
| Roles | Analysts produce analysis; PMs and partners carry the logic in their heads; the committee ratifies. | Analysts judge, frame and own: they set the questions, review automated analyses, supply expert input and sign the record. The decision architect sets the policies in DeOS Studio. The decision maker decides and overrides with a reason. |
| Committees | Memo read in the room; discussion of the narrative; time spent on the easy cases; the decision by seniority and mood. | The recommendation with options, conviction and explanation arrives before the meeting. The committee spends its time on the cases near the threshold and on the assumptions that move the answer. Noise is visible because every member's probability is recorded. |
| The analyst's day | Reading, collecting, modelling, formatting; a few names; a quarterly rhythm. | Reviewing rankings and triggers across the whole coverage; answering precise expert questions with a range and a confidence; investigating the cases the system flags; getting calibration feedback on their own judgment. |
| What the CIO sees | Performance by fund and by PM; a monthly risk report; what people tell them. | Policy adherence and decision quality across the firm; calibration by analysis, policy and person; hidden concentration on drivers; where the mandate drifts; which decisions are made outside the policy and why. |
| What investors are shown | A track record, a pitch book and a process description answered from memory in due diligence. | The decision record: every decision replayable with its context, probability, options and reason; calibration evidence; skill shown in the decisions and separated from luck, not only in the return series. |
| The human / AI mode | Everything is judgement only, unstructured and unrecorded, however good the people are. | Per component: Automated where volume is high and ambiguity low; Augmented where expert knowledge is scarce in the data; Judgement only where accountability, values or regulation require a person. The decision stays human. |
The gain, stated precisely. It is not the same analysis automated. It is a different and larger analysis: every name instead of a shortlist, thousands of paths instead of three, a calibrated distribution instead of a point, every regime at four scopes, counterfactuals as a routine input, experts as structured predictions, continuous updating, a memory that never forgets and the whole decision lifecycle optimized rather than one decision. The person still decides, and for the first time can be shown to have decided well.
06
The value equation in one page
8. From inefficiency to alpha to benefit to KPI to monetary value#
The whole argument on one page: each family of inefficiency, the alpha that removes it, the benefits it shows up in, the KPIs that prove it and the equation that prices it; then the value equation of each of the eleven decision archetypes.
| Inefficiency family | Alpha | Benefits | KPI | Value equation ($) |
|---|---|---|---|---|
| Governance: unwritten criteria, drift, key-person dependency, outcome-based review, no record, inconsistent mandate | Governance alpha | Decision quality, transparency, fundraising success, learning, scale | Policy adherence; records complete at decision time; dispersion between judges; breaches and overrides with reason | Decisions per year × capital per decision × Δ decision quality from consistency + due diligence hours saved × cost + Δ probability of a raise × size × margin |
| Information: commodity information, coverage, silos, stale data, look-ahead and survivorship, unseen alternatives, hidden variables, noise | Information alpha | Scale, speed, risk reduction, effectiveness | Coverage per person; shortlist hit rate; hidden concentration; lead time; point-in-time share | Δ hit rate × pursuits per year × value per hit + Δ probability of a hidden-driver loss × exposure × loss given event + value of windows caught |
| Analysis: small samples, overfitting, p-hacking, Gaussian and linear assumptions, correlation, point estimates, low dimensionality, no counterfactuals, calendar updating, regime blindness | Analysis alpha | Risk reduction, effectiveness, speed, innovation | Calibration and interval coverage; live against backtest; attributed cause; worst-regime outcome; evidence-to-update time | Δ probability of a tail loss × exposure × loss given event + Δ price or return captured × volume + decisions moved into the right window × value per window |
| Behavioural: level, pattern and occasion noise, biases, overload, miscalibrated conviction, sizing by rule of thumb, path dependency | Behavioural alpha | Decision quality, effectiveness, efficiency, learning | Brier score by person; sizing skill; timing regret; test-retest consistency; calibration improvement per cycle | Decisions per year × capital per decision × Δ decision quality from noise and bias removed + Δ sizing contribution × book + committee hours freed × cost |
Decision archetype · value equation#
Concrete decision policies are built from eleven decision archetypes. Each archetype's value is written as an equation with the client's own numbers before engineering and verified at the Prove and Realisation gates.
TYPE
Selection
Go / No-Go
DECISION ARCHETYPE
Selection
Conviction test
VALUE EQUATION → $
Δ shortlist hit rate × candidates pursued per year × value per successful pursuit + analyst hours saved per screen × screens per year × cost per hour
Δ probability of a bad pass × cases passed per year × average loss given a bad pass + Δ good cases not missed × value per case
Decision
Decision
optimization
optimization
optimization
Monitoring
Monitoring
Trigger design
Root cause
Δ loss avoided per early trigger × events per year + review hours saved × holdings × cost per hour
Δ value of policies improved on evidence × decisions per year + Δ cost of wrong lessons avoided (managers kept or fired on luck)
Inputs are collected in the decision audit and the maturity assessment (dimensions D1 to D6: decision governance, decision structure, context and data, prediction and judgment, learning and proof, scale and capability). The equations are never a promise; they are checked later against the recorded decisions.
© vencortex®. This paper describes the pain decision optimization addresses and the value it creates. It is not investment advice.