THE BROKEN CLOCK

Your retirement plan is lying to you. And AI just proved it.

There's a number hiding in your financial plan. You've never seen it. Your adviser probably doesn't even know it's there. But it's the most important digit in your entire retirement strategy, more critical than your stock allocation, your fees, or your insurance.

It's the expected return baked into your portfolio.

And it's wrong.

Not "maybe off by a percent" wrong. Not "we'll smooth it out over time" is wrong. Structurally, fundamentally, catastrophically wrong. Wrong in a way that's hardwired into the financial advice industry's DNA. Wrong in a way that's about to devastate a generation of Australians entering the most dangerous decade of their financial lives.

Here's what changed everything: In April 2026, a team at Columbia University and an AI investment firm called Altbridge released a paper that will haunt finance for years. They built something called "The Self-Driving Portfolio", a swarm of roughly 50 specialised AI agents that autonomously handle the entire investment process. These agents produce market forecasts, build portfolios using 20+ competing methods, critique each other's work, vote on outcomes, and deliver board-ready recommendations. All are governed by a human-written Investment Policy Statement.

Buried in the results was a bomb.

When the system's agents tried to estimate expected returns for US Growth equities, currently trading at 31 times earnings, with yields below risk-free rates, seven different methods spit out seven different answers. The historical equity risk premium, the backbone of most retail advice models, said: 13.3%. The valuation-implied method, anchored to what markets are actually pricing today, said: 3.2%.

The AI judge, a large language model tasked with picking the best estimate, ranked the historical method dead last. It gave it a 5% weight. The final blended estimate: 6.2%.

The AI's reasoning was brutal: "At current valuations, backward-looking return history is not a forecast. It is a record of conditions that no longer exist."

That gap, 13.3% versus 6.2%, isn't a rounding error. It's the difference between retirement and ruin.

___________________________________________________________________________________

THE GHOST IN THE MACHINE

But here's where it gets weird. The AI might still be too optimistic.

Because the AI treats valuation as a signal about fundamental value. What if current prices aren't just "expensive"? What if they're mechanically elevated by the persistent imprint of trading flows? What if that 3.2% valuation-implied return is itself a fantasy, assuming prices will mean-revert to fundamentals when, in an inelastic market, they might instead drift toward the path of least liquidity resistance?

Enter Jean-Philippe Bouchaud, a French physicist who thinks about markets as he thinks about earthquakes. His research reveals something uncomfortable: markets have memory.

When you trade, the price impact follows a square-root law, it scales with the square root of your trade size relative to daily volume. This isn't informational. It's mechanical. It arises from liquidity constraints and order flow dynamics. And critically, it doesn't decay quickly.

Empirical work by Gabaix and Koijen suggests a "GK multiplier" of roughly five: $1 of uninformed buying can inflate total market capitalisation by ~$5 over several months.

Why? Because in an inelastic market, large institutional portfolios are trapped. To maintain target allocations, they can only sell if prices rise enough. So when new buying pressure arrives, liquidity providers absorb it temporarily, then pass it along the chain, from high-frequency traders to medium-term funds to long-term holders, until the flow is fully digested. Along the way, the price itself becomes information. As physicist-turned-economist Fisher Black anticipated decades ago: when fundamental value is fuzzy, market participants revise their reservation prices toward the traded price. The entire supply-demand curve shifts.

The implication is brutal: sustained flows, from passive investing, benchmark rebalancing, corporate buybacks, systematic strategies, or superannuation contributions, can push prices up or down for extended periods without any corresponding change in fundamentals. Returns aren't just earnings growth plus valuation change. They're also flow persistence multiplied by market capacity.

_____________________________________________________________________________________

THE 35-YEAR DELUSION

This reframes everything we think we know about "historical returns."

When Australian advisers tell clients that equities have returned 10% annually, they're describing returns generated during an era of structurally positive net flows: compulsory superannuation, global allocation shifts, household savings flooding into listed markets. That history wasn't generated in a flow-neutral laboratory. It was generated during 35 years of credit expansion, declining rates, and relentless inflows.

Australia hasn't experienced a full credit cycle bottom since 1991. Thirty-five years. That's not a business cycle. That's a generational secular expansion, a prolonged arc of falling rates, expanding credit, compressing risk premiums, and rising asset prices with no parallel in modern Australian history.

The "history" underpinning most capital market assumptions wasn't generated across diverse economic regimes. It was generated inside one regime. And that regime may be ending.

But there's a second distortion: those historical returns weren't flow-neutral. The past three decades coincided with the rise of passive investing, the institutionalisation of superannuation, the growth of liability-driven investing, and the systematic rebalancing of trillion-dollar indices. These aren't background conditions. They're first-order drivers of price formation in inelastic markets.

If flow regimes shift, if contributions plateau, if rebalancing turns from buyer to seller, if passive flows become correlated in stress, "the return-generating mechanism itself changes."

The industry has been speaking in business cycle terms, the shorter oscillations within a much larger secular arc, while presenting that language to clients as though it captures the full range of outcomes their retirement capital might face.

It doesn't.

And the Investment Policy Statement, structuring your entire financial plan, is built on that same bet. Silently. Without disclosure.

__________________________________________________________________________________________

THE BEHAVIORAL TRAP

Here's where the story becomes genuinely strange. The AI system didn't just expose a data problem. It exposed a behavioural one.

Look at what the AI judge did. Confronted with a historical estimate of 13.3%, it applied a brutal markdown. That reflects valuation discipline, yes. But it also mirrors something deeply human: the instinct, at market peaks, to distrust the extrapolation of good outcomes. When conditions are extended, sophisticated actors become reflexively cautious. They lower forward return expectations, not always because the evidence demands it, but because the psychology of prolonged bull markets generates earned scepticism.

Behavioural finance has documented this extensively. At market peaks, professional forecasters consistently reduce return expectations. DALBAR studies show individual investors do something similar, but with worse timing, reducing equity exposure after drawdowns rather than before. Both patterns reflect the same dynamic: our intuitions about future returns are heavily conditioned by recent experience, and our attempts to correct are imprecise and poorly timed.

The cruel irony? This behavioural caution, even when it arrives, arrives late and fades quickly. Decades of research on regime detection, from yield curve models to leading indicators to forecast surveys, confirms that markets and forecasters don't identify regime transitions well. The shift from late-cycle expansion to contraction is almost always underestimated in magnitude and mistimed in identification. By the time consensus acknowledges a regime shift, the portfolio damage is already substantial.

The caution was real. The timing was wrong. The frameworks weren't built to respond.

But there's a second, more structural irony: even if forecasters do become cautious, their caution may be overwhelmed by mechanical flow dynamics. Bouchaud's research on "Ponzi funds", using detailed ETF flow data, shows that inflows can mechanically generate performance, which attracts more inflows, which pushes prices higher, reinforcing performance, until flows reverse. This isn't fraud. It's the natural outcome of an inelastic market with limited risk-bearing capacity.

The unwind, however, is rarely gentle. When everyone wants to do the same trade, when benchmark-driven rebalancing, risk-parity de-risking, or liability-driven selling coincides, the market's limited liquidity amplifies the move. This is precisely what happened during the quant quake of 2007. It's the lurking risk embedded in many stat-arb and market-neutral books today.

The behavioural trap: investors distrust peaks but cannot time the turn.

The flow trap: even if they could, market inelasticity means the turn may be sharper and more persistent than fundamentals alone suggest.

Most advice frameworks account for neither.

_________________________________________________________________________________________

WHAT THE MACHINE BUILT

The most instructive part of the paper isn't what the AI said about returns. It's what it actually did with a portfolio.

Operating in a late-cycle, stagflationary environment with 25–35% recession probability, the system's CIO agent constructed a final allocation that was modestly underweight equities at 44.9% versus 60% in a standard benchmark. It held an explicit 8.1% cash position. Within equities, it tilted toward international developed markets and away from US Growth, the most expensive and historically overstated asset class. Within fixed income, it is concentrated in duration rather than credit, with Intermediate and Long-Term Treasuries accounting for 23.1% of total weight.

The portfolio backtested with a maximum drawdown of 25.6% versus 34.3% for a 60/40 benchmark, at only a marginal sacrifice in Sharpe ratio. The CIO agent explicitly explained why the portfolio was constructed as it was, what assumptions drove it, and what would cause those assumptions to fail.

These are conditions whose absence in retail advice represents a structural failure of client communication.

Now compare that to the typical Growth portfolio inside an Australian retail advice model: higher equity weighting, no explicit cash buffer, US Growth at market-cap weight, static CMAs last reviewed twelve months ago in a different macro regime, no written explanation of what conditions would invalidate the assumptions.

The contrast isn't subtle.

But even the AI system could be enhanced by explicitly incorporating flow sensitivity. A truly flow-aware portfolio would:

- Stress-test for correlated flow shocks: Not just "what if equities fall 20%?" but "what if passive outflows, benchmark rebalancing, and LDI de-risking coincide?"

- Account for capacity constraints: Recognise that the ability to exit without moving the price is itself a function of market depth and flow direction

- Model impact persistence: Incorporate the GK multiplier to estimate how today's flows affect prices over the next 3–12 months

- Distinguish informational from mechanical moves: Use order flow data to assess whether a price move will revert (informational) or persist (mechanical)

These aren't exotic quant techniques. They're increasingly accessible via market microstructure data, ETF flow reports, and liquidity analytics.

The gap isn't technological. It's conceptual. Most advice frameworks still treat flows as noise, not a signal.

_________________________________________________________________________________________

THE KILL ZONE

The retirement risk zone, the five to ten years on either side of retirement, isn't new to financial planning. But it's never been more urgent, and it's never been more poorly served by the assumptions underpinning most advice frameworks.

This is when sequence-of-returns risk becomes maximally destructive. In accumulation, a 30% drawdown has time to recover. Returns compound in your favour. In the risk zone, the mathematics reverse. A client drawing 4–5% annually from a portfolio that immediately loses 25–30% isn't experiencing a temporary setback. They're locking in losses by selling depressed assets to fund living expenses, what researchers call "dollar cost ravaging," the cruel inverse of dollar cost averaging. The portfolio may never recover. The retirement income plan, built on historical return assumptions bearing no relationship to forward reality, collapses.

In Australia, this has a particularly sharp dimension. Under current superannuation rules, anyone receiving an account-based pension must draw down a legislated minimum each year, starting at 4% for those under 65, rising to 14% for those 95 and above. These minimums apply regardless of market performance. In fact, there are regulators who want super balances reduced to zero, which would see higher drawdown rates. A client in the risk zone who experiences severe negative returns in years one or two cannot simply reduce drawdowns to preserve capital. They're legislatively compelled to keep selling, at depressed prices, into a declining portfolio, in exactly the sequence that permanently impairs long-term outcomes.

But here's where market inelasticity compounds the catastrophe: in a flow-sensitive market, mandatory selling begets further selling pressure. When thousands of retirees simultaneously draw down from similar model portfolios, their collective redemptions become a correlated flow shock. If those portfolios hold illiquid assets, small-cap equities, credit strategies, unlisted infrastructure, the impact magnifies (interesting given the recent "we need leverage cry"). The square-root law means selling $X into a market with average daily volume $V creates a price move proportional to √(X/V). If X rises (more retirees selling) or V falls (liquidity dries up in stress), the impact accelerates.

This isn't theoretical. Bouchaud's work on ETF flows shows that concentrated, illiquid funds experience the strongest flow-performance feedback loops. Many Australian retirement portfolios—particularly those using unlisted property or private credit to "enhance yield"—sit precisely in this vulnerability zone.

Morningstar's 2025 research reduced its safe starting withdrawal rate to just 3.9%, down from the traditional 4%, specifically because of current equity valuations, bond yield structures, and inflation expectations. But even that adjustment understates the risk as it doesn't account for flow-driven impact persistence.

The historically-grounded assumptions baked into most advice frameworks haven't been revised to reflect this. The clients most exposed are the ones least able to absorb the consequences.

_____________________________________________________________________________________

THE SILENT BET

The Investment Policy Statement is supposed to be the foundational document of the client-adviser relationship. In institutional investing, it's a living governance document, updated as conditions change, stress-tested against scenarios, and reviewed by committees with current macro analysis. In retail advice, it's something closer to a compliance artifact: set at the start, reviewed annually at best, rarely interrogated for the assumptions it encodes.

The Ang, Azimbayev and Kim paper makes a point that deserves careful reading by every advice professional in Australia. In their agentic system, the IPS is the single governing document, the same document that guides human portfolio managers now constrains and directs autonomous agents. Their macro agent classifies the current environment as late-cycle with stagflationary risk, and that classification propagates through every subsequent stage. Asset class return estimates change. Portfolio construction method preferences change, in the late-cycle run, top-ranked methods relied primarily on covariance structure rather than expected return forecasts, because the system explicitly acknowledged that expected returns are less reliably estimated than correlations.

The entire architecture is regime-aware, not regime-agnostic.

Most retail advice is regime-agnostic. The IPS written for a client in 2019 is functionally identical to one written in 2026, even though the macro environment, valuation landscape, interest rate structure, and credit cycle position couldn't be more different. The return assumptions in the 2019 document were probably not far wrong for the conditions that followed. The return assumptions in a 2026 document, drawn from that same history, may be significantly wrong for the conditions now building. Seriously consider how many times in client communication you mention product names vs. regimes?

But there's a second, equally silent bet: the IPS assumes flows are neutral. It doesn't ask: What if superannuation contributions plateau? What if passive rebalancing turns from tailwind to headwind? What if correlated redemptions overwhelm liquidity in stress? It presents a static allocation as though it can be implemented at modelled prices, ignoring that implementation itself moves prices in inelastic markets. (This is such a powerful point to consider, particular when you are focused on passive investing)

When an IPS is built on CMAs derived from within-regime data and presented as a durable, forward-looking framework, it's not a neutral document. It's already made two bets, silently, without client awareness:

1. That the macro regime continues.

2. That flow regimes remain benign.

That's not financial planning. It's regime and flow extrapolation dressed in regulatory language.

__________________________________________________________________________________________

THE MACHINE THAT LEAPS

There's one further dimension that should make the advice industry deeply uncomfortable: the meta-agent.

The agentic system includes a self-improvement cycle running after each rebalancing period. The meta-agent compares the macro agent's and all asset class agents' past estimates against realised returns over a rolling three-year window, measuring regime accuracy, cross-sectional rank correlation, signal hit rates, and per-method prediction error by asset class and regime. It identifies systematic weaknesses. It researches improvements. Then it auto-modifies the relevant skill files and agent prompts. All changes are logged, with evidence and reasoning preserved.

This is qualitatively different from conventional parameter re-estimation. The meta-agent reads its own past performance, reasons about why predictions failed, and modifies both computational code and natural-language instructions. A regime-adjusted expected return method that systematically overestimates in late-cycle environments gets its confidence weighting reduced. The CMA judge skill gets updated to apply a larger valuation tilt when CAPE exceeds a threshold. The system learns, continuously and systematically, from the gap between what it predicted and what the world delivered.

Now ask the same question of the average advice practice: How often does it compare capital market assumptions against realised returns? How often does it systematically evaluate which return estimation method performed best across different macro regimes? How often does it modify its strategic asset allocation framework in response to evidence of systematic forecast error?

The answer, in most practices, is never. The assumptions are set, the frameworks are documented, and the review cycle moves on without interrogating whether the foundational inputs were right.

But even the AI system's learning loop could be enhanced by incorporating flow diagnostics. A truly adaptive framework would:

- Track the relationship between net flows and subsequent returns by asset class

- Measure the decay rate of impact post-trade to calibrate the GK multiplier locally

- Identify when flow-driven moves diverge from fundamental signals

- Adjust confidence weightings not just by regime, but by flow regime ("high passive inflow," "correlated rebalancing," "liquidity stress")

The gap between an institution that learns continuously from forecast error and flow error and one that does not is not a technology gap. It's a fiduciary gap.

_______________________________________________________________________________________

WHO PAYS THE PRICE

The clients most exposed aren't the ones with time to recover. They're the ones in the retirement risk zone—approaching or already drawing down—for whom sequence risk isn't a statistical abstraction but an immediate financial threat.

In Australia, this cohort is enormous. Superannuation balances for Australians aged 65–69 average around $430,000—held in account-based pensions subject to mandatory minimum drawdowns, in a market environment where Morningstar's research has reduced confidence in historical withdrawal rate assumptions. A severe negative return sequence in early retirement, generated by a regime shift the advice model didn't anticipate and the portfolio wasn't positioned for, can permanently impair financial outcomes in ways no subsequent recovery can fully repair.

But if markets are inelastic and flows persist, the risk isn't just that returns are lower than assumed. It's that drawdowns are deeper and more persistent because selling pressure begets further selling via mechanical impact. The legislated minimum drawdowns, intended to ensure retirees spend their savings, may inadvertently amplify the very sequence risk they cannot afford.

The advice industry has a genuine fiduciary obligation to these clients. That obligation isn't discharged by following a documented process built on assumptions never stress-tested against regime change or flow shock. The IPS isn't wrong because it uses historical data. It's wrong because it presents that data as regime-neutral and flow-neutral when it is neither, generated inside 35 years of credit expansion and structural inflows, benchmarked against business cycles embedded within a much larger and potentially ending secular arc.

____________________________________________________________________________________

THE CLOCK IS BROKEN

Here's what makes this moment different: the tools to do better are no longer theoretical.

The agentic system described in that paper runs seven methods for each asset class, blends them intelligently based on current regime and valuation context, applies structured peer review across 21 portfolio construction approaches, scores methods on estimation robustness and regime fit, and delivers a final recommendation with a full natural-language audit trail, all in minutes. The technology to construct regime-aware, valuation-conditioned, and flow-sensitive capital market assumptions at scale, to identify when historical methods should be down-weighted, and to build portfolios that explicitly position for the current macro and flow environment rather than the last one, isn't a distant aspiration.

It's operational.

For Australian advice practices, this creates both an opportunity and an obligation. The opportunity is to move beyond the single historical return estimate and build frameworks that incorporate:

- Multiple return estimation methods (historical, valuation-implied, flow-adjusted)

- Explicit regime awareness (credit cycle position, rate trajectory, inflation regime)

- Forward-looking flow diagnostics (passive allocation trends, rebalancing calendars, liquidity depth)

- Stress testing for correlated flow shocks, not just price shocks

The obligation is to stop presenting within-regime, flow-benign historical data generated inside an unusually long credit expansion and structural inflow era, as though it constitutes a full-cycle, regime-neutral, flow-neutral baseline.

Because it doesn't.

And the clients whose retirement security depends on it deserve better.

The question isn't whether the advice profession can access the analytical tools required for this shift. It can. The question is whether the profession has the intellectual honesty and structural motivation to use them, before the clients currently moving through the retirement risk zone discover, too late, that the clock their plan was built on stopped moving a long time ago.

Australia hasn't experienced a credit cycle bottom in 35 years. The generation of advisers who remember what that looks like is retiring. The frameworks built in the long expansion are still in place. The clients most exposed to the consequences of getting the next decade wrong are the ones who cannot afford to.

It's time to fix the clock.

And to acknowledge that in an inelastic market, the hands of that clock are moved not just by fundamentals, but by the persistent imprint of the flows that trade around them.

_________________________________________________________________________________________________

Primary Reference: "The Self-Driving Portfolio: Agentic Architecture for Institutional Asset Management" — Andrew Ang, Nazym Azimbayev & Andrey Kim, April 2026.

Key Microstructural Reference: Bouchaud, J-P. (2026). "Markets never forget: the lasting impression of square-root impact." Risk.net, 18 February.