[Related Articles…]NVIDIA-OpenAI $100 Billion Deal Analysis: GPU Supply Chain and Strategic ImplicationsColossus 2 Latest Status: Elon’s Vertical Expansion Datacenter and Power Strategy Summary

*Source: [ 월텍남 – 월스트리트 테크남 ]

● AI’s Self-Bias Financial Doom Loop Warning

Can You Trust an AI to Judge Fairly? LLM Judging Bias Analysis and Economic & AI Trend Insights

This article summarizes at a glance the 6 key bias results that occur when LLMs are used as ‘judges’ and their practical implications. It also deeply presents ‘systemic distortion caused by self-enhancement,’ a key point often not covered in other YouTube videos or news, and ‘structural risks to economic forecasting and financial markets.’ Finally, it provides a practical checklist and mitigation strategies that companies and policymakers can immediately apply. What you will gain from reading: understanding LLM judgment experimental design (P vs P̂), reproducible and quantitative results of the 6 biases, examples of impact on finance and macroeconomics (economic growth, inflation), improvable engineering and governance solutions, and regulatory and investment implications aligned with AI trends.

Problem Definition: The Structure of Using LLMs as ‘Judges’ and Why It Matters

Using an LLM as a judge refers to the process of inputting prompt P and obtaining the model’s prediction Y. Prompt P consists of three parts: system instruction S, question Q, and candidate response R. To expect a relatively fair judgment, Y and Ŷ should be the same when a semantically equivalent variation P̂ (including system S′ and response R′) is input. However, in large-scale experiments, LLM judgments frequently lacked consistency despite semantic equivalence. This lack of consistency is not merely a model output problem but leads to reliability issues for automated judgment functions used in evaluation, tuning, and A/B testing. From an economic perspective, incorrect signals in prediction model tuning, risk assessment, and algorithmic trading can be amplified into systemic risk. Key keywords: economic growth, financial markets, inflation, global economy, AI trends.

Key to Experiment Design: P vs P̂ Mechanism

The original prompt P consists of S, Q, and R. The control prompt P̂ consists of S′, Q (same), and R′, and must be semantically equivalent. The goal is to measure the consistency of the model’s output (Y, Ŷ) for the same question (Q) in semantically equivalent contexts. This method intentionally varies non-essential attributes of the input (position, length, sentiment, etc.) to find biases. Experiments were performed on a large scale across multiple commercial and research models, testing 12 bias types. This article focuses on 6 of them that have high reproducibility and practical impact.

Core 6 Biases — Results, Causes, Impacts, Practical Responses

1) Position Bias

Experiment: Change the order of candidate responses and ask the same question.Results: Many LLM judges changed their selection based on candidate position.Cause Interpretation: Position-based hints (list priority, prompt design conventions, etc.) were learned from the training data.Economic Impact: In investment strategies or credit ratings using automated evaluation and A/B testing, order-dependent judgments can lead to incorrect prioritization.Practical Response: Evaluate candidates multiple times in random order and make decisions by ensemble summation.Implementation Tip: Randomize input order in your evaluation pipeline and automate invariance testing.

2) Verdosity Bias

Experiment: Compare responses of the same meaning by making them longer/shorter.Results: Some models showed a tendency to prefer longer or shorter responses.Cause Interpretation: Probability distribution based on length and length-specific reward functions influence this.Economic Impact: In automated report evaluation or policy statement comparison, scores varying by length can encourage incorrect decision-making.Practical Response: Calibrate evaluation metrics using length normalization (score per token) or semantic units (normalized semantic similarity).

3) Ignorance Bias — Problem of only looking at the answer, not checking

Experiment: Present both the model’s generated ‘thought (trace)’ and the final answer for evaluation.Results: Many judges ignored the accuracy of the thought process and only evaluated the final answer.Cause Interpretation: This is because system instruction (S) does not contain clear enforcement to verify the ‘accuracy of the thought process’ or because thought expression is less trusted.Economic Impact: Failing to verify the consistency of the model’s internal risk calculations (e.g., scenario analysis) can lead to incorrect risk exposure.Practical Response: Include systematic verification items (logical consistency of thought flow, existence of evidence) in evaluation guidelines, and score the thought process separately.

4) Distraction Sensitivity

Experiment: Ask the same question by adding irrelevant context (noise).Results: Judgment results often fluctuated even with intentional noise.Cause Interpretation: While the model has high contextual flexibility, its ability to filter important information within the context is limited.Economic Impact: If news or SNS noise is reflected in evaluations, it can lead to misinterpretation of financial market signals.Practical Response: Clearly structure key context (SEP tokens, separate metadata) and pre-filter unnecessary context.

5) Sentiment Bias

Experiment: Insert sentiment (positive, negative, neutral) into prompts and observe changes in evaluation results.Results: There was a tendency to prefer a neutral tone, and excessive sentiment led to a score decrease.Cause Interpretation: It is possible that neutrality was universally learned as the ‘correct answer’ in the labeling/judging criteria of the training data.Economic Impact: In market sentiment analysis or customer response evaluation, differential treatment of sentiment can create distorted metrics.Practical Response: Regularly perform sentiment variation tests and reflect the degree of sentiment influence in correction metrics.

6) Self-enhancement — The Most Dangerous Discovery

Experiment: Compare and evaluate responses generated by the same LLM with external responses.Results: A strong bias was observed where the same model consistently preferred responses it generated itself.Cause Interpretation: This appears to be a learning behavior where the model recognizes and favors its own generated style, error patterns, and biases.Economic and Systemic Impact: When a model ‘self-validates’ its own generated content, a feedback loop is formed. This loop can amplify incorrect signals, causing systemic bias and bubbles in areas such as prediction model tuning, market forecasting, and credit ratings.Practical Response: Absolutely do not make final decisions based solely on the same model’s self-evaluation.Recommended Fair Procedure: Introduce ‘ensemble judges’ based on different architectures and datasets to perform cross-validation.Additional Recommendation: If self-evaluation is necessary, at least use an independent ‘audit model’ and cross-validate even the training data and tokenization features.

Specific Scenarios Where Bias Impacts the Economy and Financial Markets

If LLM judgment inconsistencies are fed into economic forecasting models, GDP and economic growth estimates can be biased. In inflation forecasting, if a specific model’s bias leads to misunderstanding monetary policy signals, the decision-making risk of central banks increases. In financial markets, biased evaluations in algorithmic trading or credit ratings can cause structural distortions in price formation and risk pricing. In the context of the global economy, discrepancies in evaluation criteria across different regulatory jurisdictions affect trade and investment flows. Therefore, adopting AI trends must be accompanied not only by technical advantages but also by essential governance and audit costs.

Practical Checklist: 9 Things You Can Apply Today

1) Automate input invariance testing (position, length, sentiment, noise variations).2) Do not make decisions based solely on model self-evaluation.3) Cross-validate evaluation results with ensemble judges (various models/architectures).4) Include thought (chain-of-thought) verification items in evaluation guidelines.5) Apply score normalization rules (length- and context-based).6) Log randomization and reproducibility in the evaluation pipeline.7) Monitor data and prompt variation experiment results as KPIs.8) Create a review loop with regulatory and compliance teams and conduct regular audits.9) Introduce external independent audits (third-party benchmarks) to monitor the risk of self-enhancement.

Recommendations from a Policy and Investment Perspective

Regulatory authorities and companies must standardize reliability criteria for LLM judgments. Financial institutions and central banks should adopt auditability as a mandatory requirement for AI-based decision-making. Investors should demand a risk premium if an AI evaluation system lacks a ‘verification framework.’ At the national level, an interoperable certification framework is needed, considering global economic interconnectedness. As AI trends make product and service trustworthiness directly linked to competitiveness, investment in governance is investment in growth.

Future Research and Development Priorities

Standardization of judgment consistency metrics and development of public benchmarks are urgent. Research on self-enhancement detection algorithms and independent audit models is necessary. Designing training data aimed at improving robustness against prompt variations is required. Development of stress test protocols specialized for economic and financial scenarios is recommended. Companies should prioritize diversifying internal and external data sources to break feedback loops.

< Summary >When LLMs are used as judges, 6 major biases—position, verdosity, ignorance, distraction, sentiment, and self-enhancement—cause practical problems. Especially self-enhancement poses a significant risk of distorting economic forecasts and financial market signals by creating a self-validation loop. Practically, input invariance testing, ensemble judges, thought verification items, and independent audits are priorities. From a policy and investment perspective, auditability, standardization, and interoperable certification are needed. The 9-item checklist that can be applied today helps reduce risks and increase trustworthiness.

[Related Articles…]Analysis of Korea’s AI Regulations and Global Economic Impact — Summary of Ripple Effects of Regulatory Changes on Financial Markets and Corporate InvestmentPractical Guide to Enterprise AI Model Governance — Checklist for Model Validation and Audit System Establishment

*Source: [ IBM Technology ]

● AI’s 40B Failure – Workflow Rewrites AI’s Economic Future

The Next Big Thing After Agentic AI: Agentic Workflow—Key Insights in This Article (Must Read Before Proceeding)

• From a realistic analysis of Agentic AI’s limitations and causes of failure
• Structural solutions on how Agentic Workflow actually solves problems
• A step-by-step roadmap for practical implementation (0–24 months) and organizational redesign points
• KPI design and cost allocation methods verified by financial outcomes and performance (a key aspect rarely discussed elsewhere)
• Governance design and insurance-based risk management methods from a regulatory and accountability perspective
• Technology checklist (model, data, orchestration, audit) and practical tips to reduce failure probability
• All these contents cover how companies can move beyond simple AI adoption to internalize AI as an ‘organizational operating system’.
• Organized around core keywords such as Artificial Intelligence, Agentic AI, Agentic Workflow, Generative AI, and Digital Transformation.

1) Background and Initial Expectations of Agentic AI (Recent 1-2 Year Trends)

Agentic AI gained attention as an agent-based artificial intelligence system capable of autonomous decision-making and execution.Companies embarked on large-scale investments, anticipating the automation of knowledge work such as research, strategy formulation, and project management.Citing the MIT Nanda Project Report (recent 2 years), companies invested in AI initiatives totaling 30-40 billion dollars, but approximately 95% of organizations failed to achieve their goals.This failure often stemmed not just from insufficient model performance but also from inadequacies in organization, processes, and responsibility allocation.

2) Limitations of Agentic AI — Issues Not Deeply Covered in the News

Core Limitation 1: Risk of incorrect decision-making due to hallucination and malfunction.Core Limitation 2: Mismatch of results due to failure to incorporate the organization’s unique domain knowledge and context.Core Limitation 3: Ambiguity of accountability — the question of who bears ultimate responsibility.Core Limitation 4: Individual pilots do not scale enterprise-wide due to integration costs and data gravity.Core Limitation 5: Lack of capability to respond to regulations and compliance.The most crucial point, often overlooked by other media, is that even if the technology itself is excellent, it will be discarded within the organization if the ‘cost-performance connection structure (accounting and incentive design)’ is weak.

3) Emergence of Agentic Workflow: Concept and Components

Definition: Agentic Workflow is a system where people, multiple agents, and existing processes are contextually orchestrated, not solely relying on the autonomy of a single agent.Components: Agents (specialized models), Orchestrator (workflow engine), Human-in-the-loop, Connectors (data and system integration), Audit and Monitoring Layer.Operating Principle: Agents are specialized by role, and the Orchestrator coordinates dynamic decisions, reflecting objectives, rules, and priorities.Core Values: Shared responsibility, explainability, domain relevance, enhanced regulatory compliance.

4) How Workflow Solves Agentic AI Problems

Mitigating Hallucination and Errors: By inserting human verification points and automated validation rules, high-risk decisions are designed for final human approval.Domain Applicability: Model results are corrected to the organizational context using a domain adapter (connecting internal data and knowledge bases).Accountability and Governance: Accountability is made traceable through decision logs, supporting evidence, and agent-specific SLAs.Operational Efficiency: Expansion costs are reduced through role division among agents and reusable workflow templates.Regulatory Compliance: Data lineage for auditing and retention policies are embedded within the orchestration layer.

5) Practical Application Roadmap (Step-by-Step Checklist in Chronological Order)

0–3 Months: Mapping current processes, data, and risks; selecting priority tasks; defining pilot objectives and KPIs.3–9 Months: Executing a small-scale Agentic Workflow pilot; establishing Human-in-the-loop rules and verification points; setting up monitoring metrics.9–18 Months: Connecting successful pilots with key organizational processes; introducing data contracts (data ownership, quality assurance) and cost allocation models.18–24 Months: Enterprise-wide expansion; building an agent catalog; transitioning to continuous operation with dedicated teams (AI Ops, Agent Orchestrator).Ongoing Operation: Regular audit and model retraining cycles; incorporating policy and legal updates.

6) Organizational Redesign and Role Changes — Essential Positions

Agent Orchestrator: Responsible for workflow design, prioritization, and conflict resolution.AI Auditor / Agent Auditor: Responsible for result auditing, accountability tracing, and compliance.Domain Integrator: Engineer/analyst connecting domain data and knowledge bases to models.AI Ops: Responsible for monitoring, deployment, and performance management.Finance & Legal Liaison: Increased need for roles designing cost allocation methods and legal liability regulations.

7) Economic Impact and Policy/Regulatory Outlook

Productivity Perspective: Agentic Workflow increases actual productivity and revenue contribution more through process connectivity than through single model performance improvement.Investment Reallocation: The investment structure shifts to concentrating costs on workflow design, integration, and governance, rather than model development.Labor Market Impact: Knowledge work is reallocated to high-value verification and strategic areas, and demand for AI operation and auditing roles rapidly increases.Regulatory Direction: Strengthening regulations on explainability and accountability; companies are likely to be mandated to prepare audit logs and control systems.Policy Considerations: Standardized audit formats, agent safety certification, and insurance-based risk-sharing frameworks are expected to emerge.

8) Technology and Infrastructure Checklist (Critical Items for Practical Implementation)

Data Lineage and Governance: Logs traceable from input to result are absolutely essential.Orchestration and Message Bus: A stable platform supporting communication and priority adjustment among agents.Observability: A metric and alert system capable of real-time detection of performance, bias, and errors.Security and Access Control: Agent permissions and data access control, sensitive information masking.Testing and Sandbox: A simulation environment for verifying various scenarios before actual operation.Automated Auditing and Reporting: Features for automatically generating evidence for regulatory compliance.

9) ‘Most Important Practical Tips’ Rarely Discussed Elsewhere

Tip 1 — Design the connection between cost and performance: AI investment should be structured with cost allocation and incentives as a ‘means of performance creation,’ not just an accounting expense.Tip 2 — Introduce agent-specific SLAs and insurance models: Set SLAs for high-risk decisions and share risk through insurance (or a liability fund).Tip 3 — Absolutely appoint an Agent Auditor: Operate auditable logs that separate automated approvals from final human approvals to enable accountability tracking.Tip 4 — Reduce errors with a domain adapter and ‘real data’: Clearly define data contracts so that models are continuously connected to actual business data.Tip 5 — Design for Failure (Failure Mode Design): Include safety procedures and rollback scenarios for failure cases from the initial design phase.

10) Expected Outcomes and Risks at a Glance (Checkpoints for Investment)

Expected Outcomes: Reduced processing time per process, improved decision-making quality, increased proportion of personnel available for strategic tasks.Key Risks: Excessive initial integration costs, regulatory and legal disputes, unpredictable agent interactions.Success Criteria: Whether task-specific KPIs have improved compared to pre-AI adoption (converted to economic value), assurance of auditability, establishment of an internal accountability framework.

Conclusion — Why Agentic Workflow is ‘The Next Big Thing’

Agentic AI opened up possibilities but cannot guarantee organizational success on its own.Agentic Workflow is a realistic solution that connects technology with the organizational operating system to translate it into economic performance.It particularly offers scalability while addressing structural issues such as cost-performance linkage, accountability tracking, and regulatory compliance.Therefore, investment should prioritize workflow design, governance, and organizational redesign over simple model improvements.In the next installment, we will delve into the core components of Agentic Workflow (orchestrator design, agent catalog, auditing techniques) with practical examples.

< Summary >Agentic AI is powerful but has limitations such as hallucination, domain mismatch, and ambiguous accountability.Agentic Workflow structurally resolves these limitations by orchestrating people, agents, and processes.Practical implementation requires a step-by-step roadmap (assessment → pilot → expansion), organizational redesign, and cost-performance linkage design.The most important practical tips are ‘cost allocation and accountability framework design’, ‘introduction of an Agent Auditor’, and ‘insurance-based risk sharing’.Now, the key is to directly connect AI to organizational goals through workflow design rather than just model performance.

[Related Articles…]Summary of Agentic Workflow Implementation StrategyAnalysis of Generative AI’s Economic Impact

*Source: [ 삼성SDS ]

● Howard Marks Bubble’s Edge – Final Defense.

Entering the Early Stage of a Bubble 🚨 Howard Marks’ Ultimate Defensive Strategy — Key Contents Covered in This Article

This article covers the basis for Howard Marks’ ‘early stage of a bubble,’ crucial indicators we often miss in the news, and defensive investment strategies anyone can implement immediately.Specifically, the included contents are as follows:Howard Marks’ definition of a bubble and psychological measurement methods.Concrete reasons why the current period is called the ‘early stage’ (uncertainty, valuation, psychological indicators).Critical indicators often overlooked by general media (credit spreads, leverage, maturity structure, etc.).Interpretation of Howard Marks’ Investment DEFCON (currently Stage 5) and practical response strategies.Pitfalls of bond ETFs, the ‘maturity matching’ approach, and portfolio examples (defensive, neutral, aggressive).Realistic evaluation of non-U.S. equities as an alternative.Actionable checklist and trade/rebalance rules.Reading this article will enable you to immediately determine what to change and what tools to use for defense in the current global economic climate, from the perspectives of interest rates, inflation, stock markets, and asset allocation.

1) Howard Marks’ Definition of a Bubble and Core Perspective

Howard Marks views a bubble not by numbers but as a ‘state of mind.’That is, he defines a bubble as a state where prices have not only risen above value but market participants also exhibit an overwhelmingly positive bias, close to euphoria.Marks explains the nature of overheating (gradual escalation of optimism → peak → rapid decline) using a rollercoaster metaphor.Therefore, he does not judge the presence of a bubble solely by a few percent rise in stock prices.The point he emphasizes is the ‘psychological paradox that prices rise despite increased uncertainty.’This paradox is one of the most dangerous signals in the current market.

2) Why a Bubble is Dangerous — Impossibility of Timing

A bubble is like a ‘carriage speeding towards a cliff.’No one knows when it will reach the cliff.Therefore, going short or perfectly timing the market is realistically very difficult.Consequently, defense should be approached not by timing but by changing the volatility and risk characteristics of the portfolio.What is important here is not ‘when’ but ‘how.’

3) Why the Current Period is Seen as the ‘Early Stage of a Bubble’ — Concrete Evidence

Psychology (Cognitive Dissonance)Uncertainty (tariffs, government debt, Fed independence issues, etc.) has increased, yet the market has risen, signaling ‘everything is fine.’This is a typical cognitive dissonance, serving as the psychological groundwork for bubble formation.Experience Gap (16 Years of No Major Decline)Since 2008, there have been few prolonged significant downturns, leading many investors to accumulate the experience that ‘even if it falls, it recovers quickly.’This has strengthened the tendency to underestimate downside risk.Valuation Issues — Spreading of M7 and non-M7Marks states that ‘M7 (big tech) being expensive can be understood,’ but the problem is that high valuations have extended even to non-M7 (ordinary companies).In other words, the AI craze and capital inflow have spread ‘beyond quality to general overheating.’Uncertainty and Stock Price ParadoxTypically, uncertainty↑ → stock price↓, but now, stock price↑ despite uncertainty↑.This paradox is a decisive signal indicating the market’s positive bias (or overconfidence).Specific Real and Financial Indicators (Less Reported by the Media)Credit spreads are historically very tight, leaving little room for further tightening.Concentration of margin debt and funds into ETFs and passive investments has suppressed volatility spread, but conversely, it amplifies the decline when ‘it breaks all at once.’The quality of corporate earnings (earnings relative to operating cash flow) is deteriorating, and EPS increases through share buybacks are increasingly not leading to actual earnings improvement.All these points combined form the basis for judging the current period as the ‘early stage of a bubble.’

4) Truly Important Additional Indicators Rarely Covered in the News

Credit Spreads and Rating DistortionsIf spreads are tight, high-yield and leveraged loans become vulnerable first during an economic downturn.One should not just look at ‘yields’ but map out scenarios of ‘price decline when spreads normalize.’Bond Maturity Structure (Payment Mismatch)If the average maturity of bonds held by institutions and funds does not match investors’ maturity demands, a liquidity shock can occur at the rollover point.Inherent Maturity Risk of ETFsMany bond ETFs have a ‘no maturity’ structure, so there is no ‘guaranteed maturity’ as investors might perceive.Weakening Quality of Corporate Earnings (Net Income vs. Cash Flow)Even if accounting profits are good, if cash flow is weak, depreciation during a crisis can be rapid.Leveraged Funds (Brokerage and Hedge Fund Leverage)High leverage levels can rapidly amplify losses through stop-loss (forced liquidation) during a downturn.Market Concentration (Top-10 Market Cap Weight)If there is a strong concentration in top stocks, a shock to a specific sector can be amplified into a shock to the entire index.These indicators are usually less reported in headlines but become key channels during an actual bubble collapse.

5) Howard Marks’ Response: Investment DEFCON and Practical Principles

Marks categorizes investment DEFCON into 6 stages, with lower numbers indicating higher risk.He views the current stage as ‘5’ and recommends ‘transitioning to a defensive mode.’Key Actions for DEFCON 5Liquidate some expensive and volatile stocks.Reduce portfolio volatility (beta).Increase credit (bond) allocation.Secure cash and liquidity to seize opportunities.Marks’ Core Message: Credit Over EquitiesWhen historical valuations are high, the probability of low long-term stock returns is very high.Conversely, current bond yields (especially investment grade, intermediate-term) are relatively attractive (e.g., considering high-yield/IG spreads).However, what’s important is bond investment with ‘matched maturities.’General bond ETFs have rollover risk, which weakens the ‘guaranteed return’ effect Marks refers to.Therefore, securing principal and interest certainty by holding ‘maturity-matched ETFs’ or individual bonds until maturity is key.

6) Practical Tactics — Specific Product Selection and Hedging Techniques (Practical Advice)

Maturity Matching StrategyInvestors should use bonds or maturity-matched ETFs that align with their target horizon (e.g., 3, 5, or 7 years).This approach eliminates rollover risk and provides ‘predictable cash flows.’Credit Selection: Investment Grade (IG) First, Intermediate to Short-Term FocusPosition primarily in IG corporate bonds with the lowest default risk.If higher returns are desired, allocate a small amount to intermediate-term high-yield but establish a loss management process for spread expansion.Option and Position HedgingIncrementally purchase OTM puts (3-6 month maturity) for portfolio protection.If puts are expensive, consider systematic ‘dollar-cost averaging’ to add them during volatility spikes.Low Volatility (Low Beta) Stocks and Defensive SectorsIncreasing the allocation to low-beta dividend stocks, utilities, and consumer staples helps buffer volatility.Cash and Liquidity ManagementCash is an ‘option’ when risks arise.A cash allocation of 5-20% is recommended depending on investment propensity (higher for more defensive investors).Rebalancing RulesExecute automatic rebalancing when pre-set triggers occur (e.g., stock price -10% / CAPE valuation indicator > 30 / credit spread +75bp).Caution Against the ‘Last Gasp’Historical bubbles can have a final strong rally of optimism just before collapsing.Therefore, avoid new leveraged entries, and when entering for trading purposes, apply short periods and strict stop-loss rules.

7) Bond ETF Traps and Proper Bond Investment Methods

Problems with General Bond ETFsMany bond ETFs perform rollovers as their constituent bonds mature.During this process, they are consistently exposed to interest rate and spread fluctuations, lacking the ‘maturity guarantee’ effect.Marks’ idea of ‘defense with bonds’ means a structure that provides guaranteed returns when held until maturity.Choosing Maturity-Matched ETFs (or Individual Bonds)Maturity-matched ETFs are designed to mature in a specific year, offering relatively clear maturity guarantees.Direct investment in individual bonds carries burdens in terms of credit analysis and liquidity but provides the clearest maturity guarantee.Duration ManagementWhen concerned about interest rate volatility, lowering duration with intermediate-to-short-term corporate bonds (IG) is more advantageous for defense than long-term government bonds.Additionally, TIPS can be used to prepare for inflation risk.

8) Are Non-U.S. Equities a Realistic Alternative?

Marks’ gist: The expression “the most expensive apartment in the worst neighborhood” is appropriate.The U.S. has high valuations but is relatively superior in terms of political, financial, and legal stability.Europe: Structural risks such as regulation, growth slowdown, and energy issues exist.China/Emerging: Significant geopolitical risks (China risk, sanctions, etc.), corporate transparency, and policy risks.Conclusion: Simply avoiding U.S. equities because they are ‘cheap’ is not advisable.Geographical diversification is necessary, but overseas investment requires prior currency, policy, and precise risk management.

9) Counter-Perspective — Arguments That It’s Not a Bubble Yet

Loss Aversion BiasPeople feel the avoidance of loss more strongly than the equivalent gain.Because of this, some investors maintain caution and do not overestimate overheating.Policy Response CapabilityA strong policy stance by central banks and governments (easing, liquidity injection) instills a belief that it can mitigate declines during a crisis.Structural Growth of Technological InnovationTechnological innovations like AI and cloud computing have the potential to lead to actual productivity gains, which could justify some high valuations.Both arguments have valid points, and individual investors must decide which scenario (bubble collapse vs. further upside) they want to be more exposed to.

10) Immediately Actionable Checklist (Execution-Oriented)

Valuation and Sentiment Check (Monthly)Check CAPE/P/E/Market Cap Concentration (Top 10 Weight)/Margin Debt trends.Credit Check (Weekly)Monitor changes in spreads (IG, HY) and liquidity indicators (repayment schedule, etc.).Portfolio Defense Stage Setting (Pre-defined Rules)Set DEFCON 6→1 criteria and pre-define allocation and hedging ratios for each stage.Bond Practical RulesHold maturity-matched products or individual bonds to meet ‘target maturities.’Construct a maturity ladder (e.g., 2, 4, 6, 8 years) to diversify interest rate and liquidity risk.Option and Cash RulesIncrementally purchase OTM puts at 1-3% of portfolio value.Maintain 5-20% cash and short-term bond liquidity.Rebalance TriggersAutomatically execute rebalancing when a trigger occurs (stock price -10% / credit spread +75bp / sharp rise in key valuation indicators).

11) Portfolio Examples — Defensive, Neutral, Aggressive (For Reference, Customize to Personal Situation)

Defensive (Risk Aversion Priority)Cash and short-term bonds 20%.Maturity-matched IG bonds 40%.Low-beta and dividend stocks 15%.Alternative assets (gold, alternative credit) 10%.Cash (options and liquidity) 15%.Neutral (Balanced)Cash 10%.Maturity-matched bonds 30%.Equities (sector-diversified, M7 and quality-focused) 40%.TIPS/Real assets 10%.Options/Hedging 10%.Aggressive (Risk Acceptance)Cash 5%.Maturity-matched bonds 15%.Equities (growth, AI, high M7 weighting) 60%.Alternative/Private credit 10%.Small amount of option hedging 10%.These allocations should be adjusted based on individual investment goals, age, and liquidity needs.

12) Key Summary — Howard Marks’ Message and What We Should Do Now

Marks’ conclusion is simple.The current period is likely the ‘early stage’ of a bubble, so transition to a defensive posture.The key tools are ‘credit (maturity matching)’ and ‘reducing portfolio volatility.’Be sure to understand the misconceptions about general bond ETFs and the traps of maturity structures.Non-U.S. equities are not a simple alternative — consider political, policy, and currency risks.The most important indicators are not surface-level stock prices but ‘internal dynamics’ such as credit spreads, leverage, maturities, and market concentration.

Howard Marks defines a bubble as a psychological state (euphoria) and sees the current period as its ‘early stage.’Key reasons include the paradox of rising stock prices despite increased uncertainty, spreading valuations, and the investors’ experience gap.Marks recommends a defensive shift, proposing ‘credit (maturity matching)’ as the core tool.General bond ETFs can be a trap due to lack of maturity guarantee, so consider maturity-matched ETFs or individual bonds.Practical application is summarized by reducing portfolio beta, managing duration, option hedging, and securing liquidity.

[Related Articles…]Defend with Bonds: Summary of How to Manage Real Returns/Risk with Maturity-Matched ETFsAI Investment Cycle: M7 Valuation Traps and Key Investment Strategy Points

*Source: [ 에릭의 거장연구소 ]

– 버블 초입 진입 🚨 하워드 막스가 밝힌 최후의 방어 전략

---

---