CFA Institute

As Swiss psychologist Carl Jung noted, “Until you make the unconscious conscious, it will direct your life.” Our relationship with money is often driven by beliefs formed long before we entered the world of investing. Most clients cannot articulate their money beliefs because they operate beneath their awareness. Yet these beliefs are powerful, deeply rooted, and guide behavior. For example, children from households where resources were inadequate or unstable, commonly develop an underlying scarcity belief and anxiety about “never having enough.” As adult investors, that belief may surface as hyper-control over finances or an excessive focus on performance and growth — even if wealthy. Equally, another child raised in the same circumstances may develop the opposite belief: better to spend it now, because it may not be there later. The external circumstances are the same, but the internal narrative — and therefore the financial behavior — can be quite different. Many of our money beliefs are established early in life, though some emerge later through significant life experiences. An advisor shared an experience with an ultra-high-net-worth widowed client who had long exhibited patterns of extreme frugality and tight financial control. Despite two wealth management teams offering their insights, the advisor’s team uncovered that the client’s financial behaviors were driven by a deep sense of responsibility to protect their late partner’s legacy. The belief: “If I make changes, I’ll be disloyal.” With gentle probing, the advisor led a meaningful conversation that resulted in the client’s openness to change. Many of our beliefs are inherited patterns shaped by our family of origin, and while these internalized beliefs form the foundation of our financial decisions, much of our relationship with money is also influenced by the models we learn from our parents. source

Linda H Chen, CFA, Wei Huang, and George J. Jiang Many stocks shift into value or growth each year. The value premium is stronger among these “new” stocks, especially in contractions, tightening cycles, and high-uncertainty periods, driven largely by across-industry effects. source

Blitz, David, Matthias X. Hanauer, Tobias Hoogteijling, and Clint Howard. 2023. “The Term Structure of Machine Learning Alpha.” Working paper (19 July). doi:10.2139/ssrn.4474637. Brown, Tom B., Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, et al. 2020. “Language Models Are Few-Shot Learners.” In NIPS‘20: Proceedings of the 34th International Conference on Neural Information Processing Systems, 1877–901. doi:10.48550/arXiv.2005.14165. Gu, Shihao, Bryan Kelly, and Dacheng Xiu. 2020. “Empirical Asset Pricing via Machine Learning.” Review of Financial Studies 33 (5): 2223–73. doi:10.1093/rfs/hhaa009. Konstantinov, Gueorgui S., and Frank J. Fabozzi. 2025. Network Models in Finance: Expanding the Tools for Portfolio and Risk Management. Hoboken, NJ: John Wiley & Sons. Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. “Attention Is All You Need.” In NIPS’17: Proceedings of the 31st International Conference on Neural Information Processing Systems, 6000–10. source

Natural language processing in finance is redefining how institutions analyze text data, assess risk, and extract insights from markets. Quantum computing, which allows machines to explore many possibilities in parallel so certain tasks can run dramatically faster than on today’s computers, will not instantly transform finance — but that day is coming, and practitioners should plan for it, according to the author of this chapter of AI in Asset Management: Tools, Applications, and Frontiers. The author argues that quantum computing will not remake finance overnight, but firms can gain near-term value from hybrid quantum–classical methods for hard optimization and simulation while preparing for quantum-safe security. In summary, the authors suggest that practitioners experiment pragmatically now (portfolio optimization, Monte Carlo, targeted machine learning) and begin their shift to post-quantum cryptography. Firms that begin testing mixed quantum-and-classical methods will grab early wins (faster optimization and simulations) and reduce cyber risk. Reliable, large-scale quantum computers are still far off, so near-term benefits will come from practical, small-scale quantum techniques and a careful shift to new, post-quantum encryption. This chapter shows what the move to quantum means in practice and refreshes machine learning (ML) basics — supervised, unsupervised, and neural nets — behind credit scoring, fraud detection, market/risk analytics, and portfolio construction. It spotlights the workhorses: k-Nearest Neighbor (kNN) for credit and fraud calls via nearest-neighbor similarity; k-means to flag anomalies and surface anti-money-laundering (AML) patterns; and principal component analysis (PCA) to compress correlated factors for cleaner risk and smarter allocation. source

The collar restructures the cost problem. Own a stock at $100. Buy a $95 put for $2 and sell a $110 call for $2. Net cost: zero. Downside is protected below $95, while upside is capped at $110. Early 2020 example: an investor holds a stock at $185 and implements a collar with a $175 put and a $200 call for a $50 net cost. The March crash hits and the stock drops to $150. Without protection, the position is down $35 per share. With the collar, the loss is limited to $10 per share. The drawdown is contained at 5.4% instead of 18.9%. By June, the stock recovers to $195. The investor captures most of the rally, with gains capped below $200. The result is minimal crash loss and strong recovery participation, at a $50 cost versus more than $200 for puts alone. Collars work when you are genuinely willing to accept upside caps, like in these situations: Appreciated positions held long term. Substantial gains, not sold for tax or conviction reasons, with discomfort around full volatility. Trading some upside for needed downside protection is reasonable. Portfolios with natural return constraints. Endowments targeting 7% to 8% real returns don’t need unlimited upside. Capping at 12% to 15% while protecting below -8% aligns with objectives. Favorable option premiums. Volatility skew can make out-of-the-money calls expensive relative to protective puts. You’re getting paid to sell upside you don’t desperately need. Critical discipline: Be honest about the tradeoff. If you’ll be furious watching your stock rally 40% while capped at 10%, the collar is the wrong structure. Measuring Protection at the Portfolio Level Most institutional discussions focus on the P&L of the derivative itself rather than portfolio outcomes. Wrong question. If you bought puts for $20,000 that expired worthless, did you lose $20,000? Only if measured in isolation. If your portfolio gained $150,000 while those puts prevented panic-selling during volatility, protection was worth it. Right metric: Cost of protection divided by magnitude of loss prevented in scenarios where protection actually mattered. Example: a $10 million portfolio, 80% in equities, with quarterly 5% out-of-the-money puts costing $120k annually. Three-year results: Year 1: Market +12%, puts expire worthless, cost $120k Year 2: Market -18% in Q1, puts limit the loss to -7%, saving $880k that quarter. Full year -8%, puts save ~$400k, cost $120k Year 3: Market +15%, puts expire worthless, cost $120k Total cost: $360k. Losses prevented: $400k. Net benefit: $40k. But the real value wasn’t the $40k. It was staying invested through Year 2 instead of selling at the bottom, enabling capture of Year 3 recovery. That behavioral advantage often exceeds direct P&L benefit. source

Where does DL beat classic quant? DL wins out in fast pricing/risk via neural surrogates, short-horizon forecasting from order-book data (LSTM/GRU), and cost-aware hedging with reinforcement learning. How much data is needed—and can synthetic data help?  Use as much clean, labeled history as possible. Fill gaps with VAEs/GANs for scenario expansion and privacy, then validate on held-out real data. Can Greeks and risk from neural pricers be trusted? Yes, if you use differential training (prices and sensitivities), enforce no-arbitrage/monotonicity, and monitor Greek drift in production. How can we meet latency constraints in production?  Train offline; serve compact models on GPUs/CPUs (or FPGAs for ultra-low latency); cache results; and deploy as drop-in surrogates alongside current pricers. What satisfies model risk and regulators?  Model risk teams and regulators are satisfied when you ship models with built-in explainability (feature attributions, sensitivity tests), documented data lineage, active champion–challenger (challenger models) setups, proven stability across market regimes, and explicit, enforced usage limits. Does RL work live?  It can, when trained with realistic costs/liquidity and run with guardrails (position limits, kill-switches, stress triggers) plus continuous post-trade monitoring. source

To test whether variation in the earnings–price correlation has any predictive value for stock returns, we ran regressions of correlation levels against subsequent annualized returns. The R² between S&P Composite earnings and price from 1871 through 2024 is very high at 0.95. Given the strength of this long-term relationship—and the relative rarity of low-correlation periods—it is reasonable to ask whether those periods might function as buy or sell signals. In other words, does variation in the earnings–price correlation help predict future returns? I evaluated this question across multiple rolling time horizons. The resulting R² values — linking correlation levels to subsequent annualized returns — were far lower than the R² between earnings and price themselves. For the rolling 10-year and five-year windows, the R² fell close to zero, indicating virtually no predictive relationship. The rolling 50-year period showed the strongest relationship with a R2 of 0.53. source

Under realistic borrowing limits, the Sharpe ratio fails to align with investor utility. The geometric mean — and its shrinkage-based generalized version — offer superior estimates of future performance and improve mutual fund selection outcomes. source