Vertex Macro | Machines Calculate, Markets Change

Machines Calculate, Markets Change

A proprietary trading firm's most important capability is not predicting the next price change, but stopping in time when the model is no longer reliable

The most enduring fantasy in financial markets is that a sufficiently clever machine exists that can convert the uncertainty of prices into stable profit. Given enough data, a sufficiently complex model, and sufficiently fast computation, making money seems to become an engineering problem.

People who have actually managed a trading book are usually less optimistic.

A model can have a handsome historical record, a coherent economic explanation, and an enviable Sharpe ratio, and still be a poor business. It may underestimate transaction costs, conceal market risk, depend on opportunities that cannot scale, or simply be selling disaster insurance that has not yet been called. The most dangerous models often do not look dangerous. They make money steadily, until they stop.

What a proprietary trading firm therefore operates is not prediction, but skepticism.

Excellent firms do not assume that a model will remain correct forever. They build an institution that decides how much capital a model may use, under what conditions it must reduce risk, and when it should be shut down permanently. The core product of trading is not a formula, but the decision process built around the formula.

From paper to profit, a market stands in between

Many quantitative trading strategies begin with an observation.

A certain class of stocks seems to outperform another; a certain kind of order flow may foreshadow short-term price changes; the rebalancing behavior of certain institutions may create temporary price pressure. Academic research can help discover these phenomena, but a considerable distance remains between academic significance and trading profitability.

A public market anomaly must pass through independent replication, out-of-sample validation, transaction-cost estimation, portfolio construction, and actual execution before it can become a strategy that can be allocated capital. Each step may cause an originally attractive result to disappear.

Historical data are especially generous. If researchers try enough variables, periods, and parameters, they can almost always find an upward-sloping return curve. The market has no such courtesy. It requires traders to transact at real prices, pay financing and securities-borrowing costs, and compete with others who possess similar data and models.

The best research teams are therefore not the teams most skilled at discovering handsome backtests, but the teams most skilled at destroying handsome backtests.

They ask whether the result depends on a particular period, whether the data have been revised after the fact, whether the signal decayed after public publication, and whether real orders can fill near simulated prices. A model is not an answer. It is only a capital application awaiting refutation.

Profit is not the same as skill

An upward-sloping return curve can conceal many things.

It may be merely the result of a rising market, or it may come from common risk factors such as value, momentum, credit, liquidity, or volatility. It may also collect small returns continuously in exchange for occasional large losses. Such a strategy has a handsome Sharpe ratio in calm periods, yet may return years of profit in a single crisis.

A professional trading firm therefore first explains not how much money it made, but why it made money.

If returns come mainly from rising equity markets, that is market Beta, not independent Alpha. If profit comes from bearing liquidity risk, it should be treated as risk compensation. If smooth returns come from selling tail protection, low volatility does not represent low risk. If the model used overly optimistic execution prices, what is reported is only imagined profit.

Even after all known risks have been excluded, a final unease should remain: so-called Alpha may only be a risk factor that has not yet been identified.

This is not a posture of statistical humility. It is a necessary condition of capital protection. A team that cannot explain why it makes money also cannot judge when it will stop making money.

Gross profit belongs to the backtest; net profit belongs to the firm

The most common cause of death for trading strategies is not mysterious. They are simply too expensive.

The return a strategy can actually retain can be written as a simple equation:

Net Alpha = Gross Alpha - commissions - bid-ask spread - market impact - financing costs - securities-borrowing costs

Among these, market impact is the easiest to underestimate.

Commissions are usually clearly visible; market impact is hidden inside the act of trading itself. The larger the order, the longer it takes to complete the trade, the more market depth it consumes, and the more easily it reveals the trading direction. Costs therefore do not grow linearly with the scale of capital.

This produces three classes of strategy.

The first is statistically valid and economically invalid. Its predictive power is insufficient to cover the spread and impact costs.

The second can be profitable at small scale, yet cannot accommodate large amounts of capital. Once the position expands, the trading itself destroys the opportunity.

The third has a sufficient margin of safety. Even if execution deteriorates, financing prices rise, or the signal decays slightly, net expected return remains positive.

Only the third class of strategy is worth operating as a long-term business.

A useful fragility indicator is the share of expected gross profit consumed by expected transaction costs. The higher the ratio, the less the model needs to fail completely. A slight decline in predictive power may be enough to turn net profit into a loss.

Transaction cost is not an accounting adjustment made after the model is finished. It is part of the model's definition.

Too much money is also a form of poverty

Traditional asset management treats scale as success. Proprietary trading should be more cautious.

Every strategy has capacity. As more capital is committed, the set of tradable names shrinks, the participation rate rises, the time required to build a position lengthens, holdings become more concentrated, and exit becomes more difficult. A short-horizon signal may even disappear before the order is completed.

What truly matters is not the strategy's past average return as a whole, but how much net return the next unit of capital can create:

Marginal net Alpha = marginal gross Alpha - marginal transaction costs - marginal financing costs

When marginal net Alpha falls to zero, the strategy has already reached economic capacity. Continuing to add capital may raise total revenue, yet it will lower capital efficiency and damage the performance of existing positions.

This also explains the cultural conflict between asset-management firms and proprietary trading desks.

A firm that charges on assets under management has a natural incentive to expand. More assets mean more fees, even if the return created by each unit of capital is declining. A proprietary trading desk must pay for its balance sheet. Capital that cannot produce net profit is not an achievement. It is a burden.

A firm with genuine capacity discipline must be able to refuse additional capital when a strategy is performing well, withdraw capital when marginal return is declining, and tolerate idle cash when opportunities are lacking.

In finance, refusing money sometimes proves capability more than raising it.

A risk model cannot measure the risk of the model

Trading firms usually have a large number of limits.

They constrain capital usage, leverage, volatility, VaR, single-security exposure, industry concentration, liquidity, and interest-rate sensitivity. These metrics can estimate how much the portfolio may lose if the market moves adversely.

But they cannot fully answer a more troublesome question: what happens if the model that produced the positions no longer describes the market?

Market risk assumes that the strategy remains valid and has merely encountered a temporary adverse price move. Model risk means that the distribution once used to estimate return and risk has already changed.

Trading rules may change, the original other side of the trade may disappear, more competitors may copy the same signal, financing conditions may deteriorate permanently, and data definitions may also change. A model that once reflected a structural behavior may gradually become a souvenir of historical accident.

A strategy must therefore set its loss rules while it is still profitable.

Normal operation, warning, de-leveraging, pause, and permanent retirement should be states defined in advance, not language chosen after a drawdown has already occurred. A decline in realized hit rate, deterioration in the information coefficient, deviation in fill quality, and a rising cost burden can all serve as diagnostic indicators of model health.

Otherwise, traders will readily interpret every loss as temporary volatility. The model will be given repeated chances to "observe a little longer," until a research error that could have been controlled becomes an uncontrollable capital loss.

Drawdowns speak; unfortunately, many people look only at the number

The Sharpe ratio is useful, and it is also easily abused.

For market-neutral strategies whose returns are relatively symmetric and whose tail risk is limited, it can measure the excess return earned per unit of volatility. For traditional long-only funds that operate relative to a benchmark, the information ratio is usually more appropriate. Different strategies require different performance languages.

But volatility cannot fully describe the path of returns.

Two strategies can have the same annualized return and volatility, one making money steadily, the other enduring a long losing stretch and then recovering through a few rebounds. Looking only at the Sharpe ratio, they appear similar; to a capital provider, they are entirely different.

Maximum drawdown, drawdown duration, the longest recovery period, the number of consecutive losses, return autocorrelation, and tail losses all provide additional information.

Persistent losses deserve particular attention. They may mean that the strategy is repeatedly building the same wrong position, that the market regime has already changed, or that the risk model has omitted a common driver. A drawdown is therefore not merely a performance outcome. It is a diagnostic report sent by the model.

The problem is that managers usually read good news like statisticians, and then suddenly become philosophers when they face bad news.

Seeking Alpha is also seeking who is willing to pay

Every trade has another side. If a strategy can persistently earn excess return, it should explain who is bearing the corresponding cost, and why the other party continues to trade.

Reducing the answer to "other people are irrational" is usually not enough.

Durable opportunities are more likely to come from institutional constraints. Index funds must track a benchmark, pension funds need to rebalance on a schedule, banks are constrained by capital rules, market makers need to transfer risk quickly, and portfolio managers may stay close to the benchmark in order to control career risk.

These participants do not necessarily misunderstand prices. They may know that the terms of the trade are not ideal, yet have no choice but to trade.

The most reliable Alpha therefore often has several characteristics: the other side's behavior is driven by institutions, the institutions will not disappear quickly, arbitrageurs must commit capital and infrastructure, the capacity of the opportunity is limited, and expected return is sufficient to compensate for liquidity and model risk.

Market research should analyze not only prices, but also the objectives, constraints, and evaluation systems of other participants.

Prices tell the trader what happened. Institutions explain why it will happen again.

Five bad ideas do not add up to a good portfolio

Diversification is one of the most respected words in finance, and also one of the most easily abused.

Two strategies may trade different assets, use different models, and show low historical correlation, yet lose money at the same time in a crisis. The reason is that they may jointly depend on ample liquidity, cheap financing, stable volatility, or a persistent appetite for risk.

Correlation in calm periods tells people how prices usually change. Correlation in crisis periods tells people who will be forced to sell at the same time.

Each additional strategy increases not only potential return, but also parameter error, data-mining risk, operational complexity, and exposure to unknown factors. Placing several poorly estimated models together may produce a false sense of security precise to several decimal places.

Rather than accumulating a thin advantage across several unfamiliar domains, a trading firm is usually better served by deepening a strategy it already understands. It can expand the set of tradable names, improve execution, raise predictive strength, optimize the position function, and add genuinely independent bets.

The number of trades is not equal to the number of independent bets. A hundred orders driven by the same risk factor may still, in economic terms, be a single trade.

The purpose of diversification is not to increase the number of strategy names, but to reduce the probability of shared failure.

Fast money needs machines; slow money needs patience

Short-horizon and long-horizon strategies differ by more than holding period. They require two different firms.

Short-horizon Alpha often comes from temporary liquidity demand, order imbalance, price pressure created by forced trading, and brief delays as information enters the market. Because betting opportunities are numerous, short-cycle strategies may have higher Sharpe ratios. But the capital each opportunity can accommodate is usually smaller.

In short-horizon trading, execution is part of the model.

Latency, fill rate, cancel rate, slippage, queue position, market-participation rate, and post-trade price change may all determine whether a signal can be converted into profit. A theoretically excellent signal that cannot be filled at a reasonable price is not a strategy. It is only an opinion.

Short-cycle risk management must also be faster. The team needs real-time monitoring of positions, execution, and signal decay, and must be able to reduce or stop trading automatically.

Long-horizon strategies face another set of problems.

They usually look for valuation dislocations, fundamental mispricings, mismatched risk premia, and slow diffusion of information. Transaction costs still matter, but statistical inference is more difficult. Because independent samples are fewer, researchers more easily explain too short a history with too many variables.

Long-horizon strategies require stricter out-of-sample tests, parameter-stability analysis, and portfolio risk control. Because rebalancing is slower, an unexpected industry, credit, or liquidity exposure may persist for months.

Long-horizon strategies also have a special capacity constraint: the human capacity to endure pain.

If capital providers cannot tolerate a long drawdown, even a model that is ultimately correct may be forced to close before recovery. The investment horizon, the capital horizon, and the evaluation horizon must match one another. Managing a long-horizon strategy with daily P\&L usually turns patience into a slogan and the low point into a selling point.

A bonus is also a derivative

A firm's real culture is not written on the wall. It is written in the compensation contract.

Traditional asset-management firms usually charge on assets under management. This encourages the institution to accept more capital, issue products that are easier to sell, and avoid closing subscriptions. Even if the portfolio manager sincerely hopes to raise client returns, the business model may push the firm in the opposite direction.

Hedge funds improve part of the incentive through a performance allocation, but management fees and firm valuation still remain tied to asset scale.

A proprietary trader's compensation usually depends more directly on trading profit. This looks purer, yet it produces another risk. If the trader shares the upside and does not bear future losses, he has in effect received an option paid for by the firm.

This structure encourages traders to take tail risks that are not easily discovered immediately.

The solution is not to abolish profit rewards, but to delay payment. Bonuses can be deferred, a portion of profit can be held as a reserve against future drawdowns, and the firm can also set clawback provisions. Performance should be measured by multi-year risk-adjusted profit, not by the highest P\&L of a single year.

A good compensation system should put traders, shareholders, and the risk-management department on the same time scale.

If the reward is calculated on a one-year basis, while the risk may be realized in the fifth year, then the contract itself is a defective model.

Closing a winner is harder than closing a loser

Most firms know how to start a strategy and are not skilled at ending one.

A once-successful model accumulates its own political capital. It has created profit, it is bound to the reputation of a star trader, and data, systems, and teams have been built around it. To admit that it has already failed is not only a statistical judgment. It also means writing down an organizational investment.

Model retirement must therefore have a clear governance structure.

Who can request a reduction in risk? Who has the authority to pause trading? Can the risk department overrule the model owner? What evidence is required to bring it back online? After it is closed, how is capital reallocated?

If these questions have not been answered in advance, the model owner usually becomes the last defender of his own model.

The ability to close a once-profitable strategy is an important mark of a mature trading culture. It means the firm is loyal to capital, not to historical performance, internal status, or a trader's self-assessment.

The most important model is the model that admits error

Alpha depreciates.

Competitors will discover similar opportunities, more capital will enter, transaction costs will rise, institutions will change, and the original other side of the trade may also disappear. Sometimes the so-called edge was only statistical luck from the beginning.

The core capability of a proprietary trading firm is therefore not to be forever correct, but to shorten the time between the occurrence of an error and the admission of the error.

It must be able to discover opportunities, distinguish false Alpha, control costs, constrain capacity, monitor decay, and withdraw capital when the evidence changes. Historical return is only the starting point of an investigation, not the final proof of capability.

What a professional capital allocator should truly ask is not how much money a model made in the past, but why it made money, who paid for the profit, how much capital it can accommodate, and what evidence would force the firm to stop using it.

The market will not pay a firm permanently because the firm is clever.

It will occasionally reward those who discovered the opportunity earlier, and it will not hesitate to punish those who discovered too late that they were already wrong. Genuine competitive advantage usually lies not in how much faster one predicts the market, but in how much earlier one admits error.