Product × Finance × AI: Financial Products in the Age of AIPart Two Changing one parameter中文
Chapter 19

What did not changeTime, risk and irreversibility

The scribe makes research, checking and execution almost free, lets every household afford a personal agent, and can even ask people to wait when they panic. Yet there are a few things it cannot change at all. Grain is still harvested once a year, and saving enough grain for old age still takes decades. Diversification removes only the bad luck of a single household; a drought that strikes the whole village still comes, and only those who bear it are compensated. Some losses, once they occur, can never be undone. Any constraint that survives the rerun is not one this technology can solve, and these are precisely the problems finance was first meant to handle.

1. Grain is still harvested once a year

The scribe cannot make the harvest come early, nor change when a lifetime's income arrives. People earn more when young and nothing when old, whoever keeps the accounts. Saving enough grain for old age still means starting young and saving year by year. Compounding needs time; what the scribe can save is fees and mistakes, and it cannot shrink thirty years into three.

Consider what the scribe can do here. It can work out a household's income and spending over a lifetime, set how much to save each year according to the lifetime level, and carry this out on time every year. The reason for smoothing consumption has not changed, nor has the time smoothing takes. Only the premise that "people act on their plans" has changed: it is easier to satisfy than before.

Today's counterpart. The outcome of retirement saving depends largely on how early one starts and how many years one keeps it up.

The scribe can save fees, not years

Save 12% of income each year, with ship shares returning 6% a year on average. Drag the number of years of saving, compare paying 1.5% in fees with paying 0.2%, and see how many years of income the savings amount to.

Fees 1.5%Fees 0.2% (scribe)

Illustrative: 12% of income is deposited at the start of each year and compounds at the return net of fees. The vertical axis is in years of income.

2. Ships can still sink

The scribe can spread each household's ship shares across every ship in the village at almost no cost. Diversification removes the bad luck that belongs to one ship alone: this ship strikes a reef, that one comes home fully laden, and together the result is steady. But however well spread the holdings are, diversification cannot remove a storm that strikes the whole village's ships at once. The more ships one holds, the closer the portfolio's fluctuation comes to that of the village fleet as a whole, but it never falls below it.

Statman (1987) calculated portfolios made up of randomly selected stocks: once trading costs are taken into account, a well-diversified portfolio needs at least thirty to forty stocks, and however many stocks it holds, its risk never falls below that of the market as a whole.

Today's counterpart. An index fund that holds every stock in proportion to its market share has already removed almost all individual risk, yet when the market as a whole falls, it falls too.

What diversification removes, and what it cannot

Each ship's shares fluctuate 35% a year, of which the part that moves with the whole village fleet corresponds to 16%. Drag the number of ships held and see how far the portfolio's fluctuation falls.

Illustrative: holding n ships in equal weights, portfolio fluctuation = √(16%² + (35%² − 16%²) ÷ n). The dashed line is the systematic risk that cannot be diversified away. Statman (1987) found a similar shape using actual stock data.

3. Only bearing risk earns compensation

Since individual risk can be diversified away almost for free, the market will not pay anyone for bearing it: whoever fails to diversify bears, for nothing, a risk others need not bear. Only the part that cannot be diversified away, the part that rises and falls with the whole village fleet, needs compensation. The more a ship share moves with the fleet, the higher the return its holders demand.

At this point it becomes clear what the scribe cannot do. It cannot create a ship share with a high return that does not move with the whole village. If one existed, every scribe would buy it, its price would be bid up at once, and its return would fall accordingly. An agent that promises "high returns with no risk" is either hiding risk or has made an error in its calculations.

The model's empirical record is not good. Fama and French (2004) note that its empirical performance is poor enough to invalidate many of its applications; the problem may lie in its simplifying assumptions or in the difficulty of testing it. But the reasoning that "only risk that cannot be diversified away needs compensation, and diversifiable risk earns none" does not depend on the specific form of the model.

Today's counterpart. The guiding opinions on asset management issued in 2018 by the People's Bank of China and other authorities prohibit financial institutions from promising principal protection or guaranteed returns in their asset management business.

Only risk that cannot be diversified earns compensation

The horizontal axis is how much a ship share moves with the whole village fleet; the vertical axis is its expected return. Points on the line are priced by the market. Pick the point with "high return without moving with the fleet" and see what happens.

Illustrative: the risk-free return is 2%, and a ship share that moves one-for-one with the fleet has an expected return of 7% (Sharpe, 1964). Individual risk can be diversified away and does not affect a point's position on the line.

4. Some losses cannot be undone

One kind of loss differs from all others: once it occurs, there is no chance to make it good. An old man draws grain from his savings every year to live on. If a sharp fall happens to come in the first few years of his retirement, the ship shares he sells at the bottom cannot wait for prices to recover; with the same run of years in reverse order, his old age is far more comfortable. A household that borrows to buy ship shares and is forced to sell when they fall is no longer in the game when prices later return to where they were.

Consider what this means for decisions. If a decision can be withdrawn at any time, one may as well act first and see. If it cannot be withdrawn once made, and news about the future keeps arriving, then keeping options open has value in itself: do nothing yet, or do a little less, and decide when there is more information. The scribe can make every step faster, but it cannot make an irreversible decision reversible.

Today's counterpart. For withdrawals in retirement, the same average return can lead to very different outcomes depending on the order of good and bad years. When a margin account is forcibly liquidated in a sharp price fall, the loss is fixed even if prices recover afterwards.

The same years, in a different order

An old man retires with 100 portions of savings and draws 6 portions a year to live on. The twenty years of returns are the same, only in a different order: in one, the sharp falls come in the first years of retirement; in the other, in the last years.

Illustrative: the annual returns over the twenty years are the same set of numbers, only in reverse order; without withdrawals, the two orders give exactly the same result after twenty years. The vertical axis is in portions.

5. What changed and what did not: Civilization Simulator, version 19

The simulator unlocks the last item of Part Two: "What did not change". Forty households save for thirty years, and the simulation runs once under each of eight different runs of years. Each run compares two cases: cognitive labour at its original price, with each household managing its own affairs and paying higher fees; and cognitive labour almost free, with every household having a personal agent that acts on its plan. See how much improvement the agent brings in each run of years, and whether the gap between different runs narrows as a result.

Civilization Simulator · Chapter 19: what changed and what did not

Forty households save for thirty years, and the simulation runs once under each of eight different runs of years. Each run compares two cases: cognitive labour at its original price, with each household managing its own affairs; and cognitive labour almost free, with every household having a personal agent that acts on its plan. Pick a run of years to see the details.

Saving and seedLendingInsuranceParameter: price of cognitive labourPersonal agentGuaranteeing agentsAnalysts for allAgents as market participantsAgents and human biasWhat did not change
Cognitive labour at original price, self-managedEveryone has a personal agent acting on their plan

Illustrative model: the runs of years are generated as in the "keeping to the plan" simulator, and below each run is the annualized return of ship shares over thirty years. When self-managed, every household saves the single rate of 12%, affected by its weight on the present, buys more after gains and sells in a panic after falls, and pays fees of 1.2% a year. With a personal agent, each household saves on time at its own suitable rate, panic selling falls to one fifth, and fees are 0.2% a year. The vertical axis is median savings in year 30, measured in years of income.

6. What this step leaves open

At this point in the rerun, time, risk and irreversible losses have not changed; what has changed is the cost of delegation, checking, information and execution. What remains are the questions a product must answer:

  1. What to promise. A household has its own goals, time horizon and capacity to bear risk. What can an agent promise it, and what can it not?
  2. The user is not present. When an agent keeps working on someone's behalf, what must be agreed in advance?
  3. The user cannot see the process. How can users understand and check what the agent has done?
  4. Agents make mistakes. How can irreversible consequences be capped, require confirmation and be compensated?
  5. Trust takes time. On what basis does a user begin to trust a new agent, and how is that trust built step by step?

What this step leaves open

Five questions left for a product to answer. Pick a question to see the constraint behind it.

Part TwoAfter the parameter changes

7. Implications for products

The easiest mistake for an agent-centred wealth product is to take the collapse in the cost of cognitive labour for the disappearance of finance's basic problems. Three implications follow. First, what a product can improve is fees, mistakes and execution; what it cannot improve is time, systematic risk and irreversible losses. It should promise users only the former. Second, a product must explain systematic risk clearly: how much the user bears, how much compensation that may bring, and what the worst run of years would look like. Third, every irreversible action must be deliberately slowed down, leave room for options, and be confirmed by the user in person; the faster the execution, the more it needs to stop at these points.

Chapter summary: what did not change

Four conclusions, each tied to one mechanism in this chapter.

TimeFees can be saved, years cannot

When income and needs arrive does not change, and neither does the number of years accumulation takes.

RiskWhat cannot be diversified must still be borne

Individual risk is diversified away almost for free; the risk the whole village shares is not reduced at all.

CompensationNo high return without bearing risk

Only risk that cannot be diversified earns compensation; an agent cannot put people above the line.

IrreversibilityThe faster the execution, the more room to leave

Some losses cannot be undone; keeping options open has value and must be deliberately preserved in the design.

References