A Tale that Explains Why The Market Has Moved Beyond the Order…
Algorithmic trading created a loop around execution. AI will create a loop around the entire investment process — and reduce the traditional OMS to infrastructure.
By Carl Carrie
At 10:17 a.m., Daniel’s portfolio changed.
Daniel was a buy-side trader. His order management system displayed the morning’s work in tidy rows: security, side, quantity, limit, benchmark, status.
Then a large cloud company increased its forecast for AI capital spending. Daniel’s OMS treated the announcement as news about one company. The market treated it as news about almost everything.
Semiconductors moved first. Then memory, optical networking, servers, cooling systems, data centers and electrical equipment. Utilities exposed to data-center demand rose. Copper futures reacted. The cloud company’s bonds weakened as investors reconsidered free cash flow and future financing.
Options delivered a different verdict. Implied volatility rose, suggesting that the announcement had changed not only the expected outcome, but the range of possible outcomes. Dealers recalculated their deltas and adjusted their hedges.
Index futures moved before many analysts had finished reading the release. AI and semiconductor ETFs began transmitting demand across baskets of underlying shares.
On Daniel’s screen, the reaction appeared as hundreds of separate orders.
It was one event.
The OMS knew everything about each order and almost nothing about the system connecting them.
The first loop was built around execution
This had happened once before.
At the beginning of the century, regulators in the United States and Europe were trying to make markets fairer, more competitive and more protective of investors.
Regulation NMS in the United States protected displayed quotations across trading venues. MiFID in Europe promoted competition among venues and strengthened best-execution obligations.
The goals were reasonable. A small investor should not receive an inferior price merely because an order arrived at the wrong exchange. Competition should reduce the power of incumbent markets. Electronic quotations should improve transparency.
The rules also helped create a market in which speed became an economic advantage.
Once the best displayed price might exist on another venue, firms needed systems capable of observing many venues and routing between them. Once liquidity fragmented, a participant connected to more markets could see and act on more information. Once prices were linked electronically, small discrepancies could be identified and traded before slower participants responded.
Regulation did not create algorithmic or high-frequency trading by itself. Faster computing, electronic exchanges, quantitative methods and cheaper connectivity were already pushing markets in that direction.
But Reg NMS and MiFID shaped the arena in which these technologies became dominant.
Rules intended to protect investors and encourage competition produced an unintended consequence: an arms race in market data, connectivity, routing and speed.
The market increasingly rewarded the fleet-of-foot.
The first generation of automation placed a computational loop around the order:
- Observe prices across venues.
- Select the best destination.
- Route, amend or cancel.
- Observe the resulting market state.
- Repeat.
The OMS still created and controlled the parent order. But execution beneath it became continuous, fragmented and algorithmic.
The machine was below the order.
Passive investing changed what was being traded
A second transformation changed the scale of the decision.
Index funds and ETFs did not simply create more orders. They made baskets, benchmarks and flows more important.
An investor could express a view on technology, volatility, credit, clean energy or artificial intelligence through one instrument. Authorized participants and market makers translated that demand into creations, redemptions and hedges across the underlying securities.
By the end of 2025, U.S. ETF assets had exceeded $13 trillion. Over the decade through 2025, index domestic equity mutual funds and ETFs received $2.9 trillion in net new cash and reinvested dividends, while actively managed domestic equity mutual funds experienced $3.4 trillion in net outflows.[1]
The market increasingly moved at portfolio scale.
One ETF flow could affect hundreds of securities. An index rebalance could create concentrated demand near the close. An options position could force a dealer to change hedges throughout the day. A futures trade could alter broad exposure before the underlying cash securities adjusted.
The economically meaningful object was no longer always the individual order.
It might be a basket, factor, hedge, portfolio transition or flow transmitted between related markets.
The market had become a graph.
The OMS remained a list.
AI creates the second loop
The first loop automated execution after the investment decision had been made.
AI creates a loop around the decision itself. Anthropic offers a useful description of an agent:
“They are typically just LLMs using tools based on environmental feedback in a loop.” [2]
In institutional investing, that apparently simple loop becomes economically powerful:
- Observe new information.
- Determine what changed.
- Update signals and uncertainty.
- Recalculate portfolio risk.
- Rerun the rebalance optimization.
- Apply deterministic controls.
- Generate, modify or cancel orders.
- Observe prices, fills and market response.
- Repeat.
The language model does not replace financial mathematics.
It orchestrates it.
When the AI-capex announcement arrived, an agent read the release and compared it with prior guidance. It extracted changes from tables and footnotes, identified affected suppliers and competitors, and mapped the implications across equities, options, futures, credit and commodities.
Statistical models updated expected returns and volatility. A factor model recalculated exposures. Options models repriced nonlinear risk. A transaction-cost model estimated impact and capacity. The portfolio optimizer ran again.
Then deterministic controls tested the result.
The AI could not explain away a mandate restriction. It could not invent cash, exceed a concentration limit or trade an ineligible instrument. If the proposed rebalance was infeasible, it had to revise the plan and rerun the optimization.
This was not a chatbot attached to an OMS.
It was a controlled feedback loop around the investment process.
AI makes signal discovery cheaper — and belief more expensive
At 10:18, the simplest models saw “higher spending” and “artificial intelligence.”
They bought the theme.
More sophisticated systems asked what the spending meant.
Was it evidence of stronger demand? Or evidence that earning revenue from AI would require more capital than investors expected? Would suppliers gain pricing power? Would competitors be forced to match the investment? Would depreciation reduce future earnings?
Each interpretation implied a different portfolio and a different horizon.
Options and dealer hedging could dominate the next several minutes. ETF and index flows could shape the trading session. Supplier earnings revisions might persist for quarters. Demand for power generation could extend for years.
The announcement did not produce one signal.
It produced hundreds of micro-alphas: small, uncertain predictions whose value decayed at different speeds. AI makes generating and testing these hypotheses dramatically cheaper. It also makes generating convincing mistakes cheaper.
A research team that once tested dozens of ideas can now test thousands of features, prompts, parameter combinations, models and random seeds. Several will produce attractive historical Sharpe ratios by chance.
Get carl carrie’s stories in your inbox
Join Medium for free to get updates from this writer.
The more experiments the system performs, the less believable the winning backtest becomes.
Sharpe deflation therefore becomes an operational control.
Recent research into financial deep-reinforcement-learning strategies found that apparent outperformance varied materially across random seeds and often weakened after accounting for multiple runs and selection effects. Performance needed to be treated as a distribution across learned policies — not as one privileged backtest.[3]
Before a signal reaches the portfolio, its apparent Sharpe must survive a sequence of haircuts:
- Search deflation: How many experiments were conducted before this result appeared?
- Dependence deflation: Are several models rediscovering the same factor?
- Crowding deflation: How many competitors are likely to possess the same signal?
- Regime deflation: Is the current environment represented in the training history?
- Capacity deflation: How much capital can enter before the strategy moves its own market?
- Implementation deflation: What remains after spread, impact, hedging, financing and exit costs?
A backtested Sharpe of 1.8 might fall to 1.2 after accounting for model search, 0.9 after removing duplicated signals, 0.6 after crowding and regime uncertainty, and 0.3 after implementation at the required scale.
Only the first adjustment corresponds directly to the formally established Deflated Sharpe Ratio. The broader waterfall is an economic framework rather than one standardized statistic.
Its purpose is simple: Do not let AI confuse discovered performance with realizable performance.
The OMS could change speed. Daniel needed to change behavior.
Daniel already had an order to buy the cloud company.
The conventional execution response was to adjust the participation rate. Trade faster if the announcement strengthened the thesis. Slow down if volatility and impact had risen. Perhaps switch algorithms.
The new information raised a more important question:
Should the order still exist?
Higher investment might improve long-term growth while reducing near-term free cash flow. The better expression might be a supplier basket. It might be an options structure that benefited from greater uncertainty. It might be a relative trade between companies with different exposure to AI infrastructure.
Or the correct response might be to cancel the order and wait for ETF demand and dealer hedging to pass.
These were not changes in trading speed.
They were changes in investment behavior.
The execution algorithm asked:
How should we complete this order?
The AI investment loop asked:
Given what we now know, what should the portfolio own?
Two thousand positions changed — but not independently
By 10:21, the firm’s updated signals had changed the theoretical weights of more than 2,000 positions.
The system did not create 2,000 orders.
Buying the cloud company, semiconductor manufacturers, a technology index, an AI ETF and call options might look diversified in an OMS while concentrating the portfolio in one economic thesis.
The system first dampened the duplicated signals.
It distinguished independent information from repeated exposure to growth, momentum, technology beta, long-duration equities and short volatility. It shrank forecasts whose evidence was weak and reduced the combined weight of correlated models.
Recent portfolio research has shown how large sets of correlated assets and signals can be reorganized into uncorrelated managed portfolios — a way to reveal how many independent return sources actually exist beneath an apparently diverse signal set.[4]
The system then determined which theoretical changes were worth trading.
A model might predict 12 basis points of excess return in one security. If entering and exiting would cost 15, the signal had no economic value. If the signal would decay within an hour, slow execution could eliminate it. If trading quickly moved the market, immediate execution could do the same.
Research published in 2023 on dynamic portfolio selection combined transaction costs with signals possessing different persistence. The resulting portfolio moves smoothly toward a forward-looking target rather than blindly trading to the latest theoretical optimum.[5]
Daniel’s optimizer concluded that only 600 of the 2,000 position changes were worth implementing.
Even those 600 trades were not independent orders. They formed one portfolio transition.
Index futures reduced immediate beta. Sector ETFs controlled broad technology exposure. Options preserved desired convexity. Individual securities were traded according to liquidity and signal persistence.
As fills arrived, hedge ratios changed. As prices moved, option deltas changed. As ETF flows reached the underlying shares, liquidity estimates changed.
Each change supplied new feedback.
The optimization ran again. The controls ran again. The orders changed again. Daniel’s OMS could display 600 orders.
It could not represent the fact that all 600 were conditional components of one continuously changing portfolio decision.
The LLM is the planner, not the alpha model
A language model can interpret an earnings release, connect entities, compare management language and plan an investigation.
It need not — and should not — perform every financial calculation itself.
An AI-native investment platform separates the roles.
A language model coordinates the work. Statistical and machine-learning models estimate returns and volatility. Factor models expose common risk. Probabilistic models represent uncertainty. Options models calculate nonlinear sensitivities. Transaction-cost models estimate capacity and impact. Optimizers balance return, risk, liquidity and constraints.
A research agent identifies possible signals.
A validation agent estimates how much of the reported Sharpe may be an artifact of model search.
A portfolio agent dampens duplicated exposures.
A liquidity agent determines whether the alpha survives implementation.
An execution agent compares the cost of trading with the cost of waiting.
A supervisory agent reduces authority when models disagree or enter an unfamiliar regime.
The order is an output of this process.
It is no longer the organizing principle.
The quiet demotion of the OMS
At the end of the day, Daniel’s OMS showed that nearly every order had been completed.
It knew the quantities, prices, brokers, and allocations.
It did not know that a single revised footnote in one earnings release had altered the meaning of hundreds of positions across equities, options, ETFs, futures, and credit. It did not know which signals were genuine, which were duplicates, or why 1,400 proposed trades had correctly never become orders at all. For those answers, Daniel had to close the window.
Above the ledger sits the real engine: an autonomous Ralph Loop paired with the unforgiving discipline of the Deflated Sharpe Ratio.
The AI does not fire a single, brittle prompt and hope for the best. It operates inside a persistent, recursive control loop — ingesting fresh market context, proposing portfolio adjustments, and slamming them directly into deterministic backpressure. This is where the Deflated Sharpe Ratio does its vital work: mathematically stripping away selection bias, penalizing trial-and-error data mining, and rejecting overfitted noise before capital is ever exposed. When a proposed thesis fails against the DSR hurdle, borrow limits, or covariance boundaries, the Ralph Loop does not collapse; it consumes the failure surface, refines its parameters, and iterates again. It self-corrects continuously until probabilistic intent is thoroughly tempered by deterministic mathematical reality.
The OMS will not die; it will simply be demoted. It will remain in the server racks to timestamp fills, enforce hard compliance checks, and keep the regulatory books clean. But the seat of alpha — interpreting context, calibrating certainty, arbitrating between competing hypotheses, and deciding when to hold fire — has already migrated into that recursive loop.
Algorithmic execution automated the plumbing. The autonomous loop is automating the intent.
The OMS was built to be the fund’s ledger. It was never meant to be its mind. Going forward, the ledger merely writes down the history; the autonomous loop authors the future.
Sources
[1] Investment Company Institute, 2026 Investment Company Fact Book.
https://www.icifactbook.org/pdf/2026-factbook.pdf
[2] Erik Schluntz and Barry Zhang, “Building Effective Agents,” Anthropic, December 2024.
https://www.anthropic.com/engineering/building-effective-agents
[3] “Unstable Gains: Multiplicity-Aware Evaluation of Financial Deep Reinforcement Learning,” Journal of Finance and Data Science, 2026.
https://doi.org/10.1016/j.jfds.2026.100205
[4] Nick Firoozye, Vincent Tan and Stefan Zohren, “Canonical Portfolios: Optimal Asset and Signal Combination,” Journal of Banking & Finance, September 2023.
https://www.sciencedirect.com/science/article/abs/pii/S0378426623001577
[5] Nick Firoozye, Vincent Tan and Stefan Zohren, “Dynamic Portfolio Selection under Transaction Costs and Signal Decay,” revised December 2023.
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4425407
[6] U.S. Securities and Exchange Commission, “The Trade-Through Rule and Locked and Crossed Markets Provisions of Regulation NMS,” June 2026.
https://www.federalregister.gov/documents/2026/06/17/2026-12163/the-trade-through-rule-and-locked-and-crossed-markets-provisions-of-regulation-nms
[7] International Organization of Securities Commissions, Artificial Intelligence in Capital Markets: Use Cases, Risks, and Challenges, 2025.
https://www.iosco.org/library/pubdocs/pdf/IOSCOPD788.pdf
[8] Financial Stability Board, The Financial Stability Implications of Artificial Intelligence, November 2024.
https://www.fsb.org/2024/11/the-financial-stability-implications-of-artificial-intelligence/
