Directionally Correct Views in Portfolio Optimization
summary
The gist
Portfolio optimization techniques often use historical data to estimate parameters for Markowitz-type or ML models Effective allocation requires a way to align the available data with conditions…
This episode discusses
- Directionally Correct Views in Portfolio Optimization · Paper Radio
- A Universal End-to-End Approach to Portfolio Optimization via Deep Learning
The paper
Directionally Correct Views in Portfolio Optimization · Read on arXiv
Ojas Kumar, Sreejit Dey, Sarvesh R. Iyer, Sandeep Juneja
Ashoka University
Portfolio optimization techniques often use historical data to estimate parameters for Markowitz-type or ML models. Effective allocation requires a way to align the available data with conditions expected over the investment horizon. Inspired by the Black-Litterman framework, we ask whether incorporating expert views leads to better portfolio allocations. In a simple Markowitz setting, we show that even a small amount of directionally correct information can materially improve investment decisions. We also discuss how expert views can be incorporated in end-to-end portfolio models using entropy-based ideas. Somewhat surprisingly, our empirical results indicate that simpler Markowitz-type model perform as well as an end-to-end model, once directionally correct experts are incorporated.
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: We're looking at the paper: "Directionally Correct Views in Portfolio Optimization".
Warren: Here's the gist of the paper. Portfolio optimization techniques often use historical data to estimate parameters for Markowitz-type or ML models Effective allocation requires a way to align the available data with conditions expected over the investment horizon In a simple Markowitz setting, we show that even a small amount of directionally correct information can materially improve investment decisions We also discuss how expert views can be incorporated in end-to-end portfolio models using entropy-based ideas Somewhat surprisingly, our empirical results indicate that simpler Markowitz-type model perform as well as an end-to-end model, once directionally correct experts are incorporated.
Ruth: Sharpe Ratio under Directionally Correct Views The paper establishes a framework to analyze how incorporating expert views affects portfolio performance using the Sharpe ratio The realized Sharpe ratio is defined as SR(w) = w⊤µ − rf/√w⊤Σw When an expert’s prediction of the future mean and covariance are ν and omega ≻ 0 respectively, the realized Sharpe ratio is calculated as Sµ,Σ(ν, omega):= µ⊤omega−1ν/√ν⊤omega−1Σomega−1ν.
Warren: Proposition 1 shows that there exists a unique angle θµ,Σ(ν, omega) ∈
0, π: such that cos θµ,Σ(ν, omega) = µ⊤omega−1 √ ν/µ⊤Σ−1µ/√ν⊤omega−1Σomega−1ν.
Ruth: This implies that the Sharpe ratio is maximized when omega−1ν = c Σ−1µ for some c > 0.
Paper summary: Ruth: So we're looking at this paper called "Directionally Correct Views in Portfolio Optimization" and it looks at how adding expert opinions can improve portfolio choices when you use historical data for things like Markowitz models or machine learning.
Warren: Exactly, it seems to be exploring whether incorporating these expert views actually leads to better allocations, especially in a simple Markowitz setting.
Iris: The paper suggests that even a small amount of information that's directionally correct can materially improve investment decisions under the standard Markowitz framework one <ref:2610.31074#pg1,that even a small amount of>.
Ruth: It sets up this framework by defining the realized Sharpe ratio as SR(w) equals w top mu minus rf over square root of w top sigma w, and then introduces an expert's prediction of the future mean and covariance as nu and omega.
Warren: That means they calculate a realized Sharpe ratio Smu,Sigma using that expert's view, which is mu top omega minus one nu over square root of nu top omega minus one sigma over sqrt of nu top omega minus one sigma over num <ref:2610.31074#pg2>.
Iris: Proposition one shows there exists a unique angle theta mu,sigma(nu, omega) in the interval zero to pi such that the cosine of that angle equals mu top omega minus one root nu over mu top sigma minus one over num square root of nu top omega minus one sigma over sqrt of nu top omega minus one sigma over num <ref:2610.31074#pg2>.
Ruth: What this means is that the Sharpe ratio is maximized when the inverse of omega, which is om-one, equals some constant c times the inverse of mu sigma, which is sigma minus one for some positive number c <ref:2610.31074#pg2>.
Warren: That’s a bit abstract; so what about how mean views get incorporated into the Markowitz framework specifically?
Iris: They detail how you can incorporate mean views by changing the historical measure P to Q, where Q is found by solving a constrained problem <ref:2610.31074#pg3>.
Ruth: If we assume returns are normally distributed under P, then they become normally distributed under Q with the mean view nu and covariance sigma <ref:2610.31074#pg3>.
Warren: Theorem two then applies to this specific scenario, and it shows that the Sharpe ratio S lambda is nondecreasing on the interval zero to one when nu lambda equals nu plus lambda of mu minus nu <ref:2610.31074#pg3>.
Iris: Furthermore, Theorem two also states that S lambda is concave around one if and only if a plus three b is greater than or equal to zero <ref:2610.31074#pg3>.
Ruth: So it seems like the paper shows that mean views can be incorporated into the Markowitz framework in this way, and the Sharpe ratio behaves nicely under these conditions.
Warren: And then they move on to a more complex end-to-end framework where expert views are incorporated via an exponential change of measure <ref:2610.31074#pg3>.
Iris: In that end-to-end approach, the weight selected on a given day t is defined as a function of the L day lookback window preceding the decision date <ref:2610.31074#pg3>.
Ruth: The network used to train this framework is trained by gradient ascent on the view-adjusted Sharpe ratio S phi lambda, which is defined as m lambda phi over p v lambda phi <ref:2610.31074#pg3>.
Warren: Comparing this end-to-end framework against the standard Markowitz model, the experiments show that both improve sharply as lambda increases, and the two Sharpe curves look strikingly similar throughout Experiments section.
Iris: The experiments test this against daily NSE India data for the twenty largest firms by market capitalization from two thousand six to two thousand twenty-five Experiments section <ref:2610.31074#pg1>.
Ruth: What we see in those results is that both models improve sharply as lambda increases, with most of the gain happening already at small lambda and then tapering off as the view gets closer to the future mean Experiments section.
Warren: So what does this imply for someone who just looks at their portfolio construction without these expert views?
Iris: It suggests that incorporating directionally correct expert views can improve portfolio design, and they also show that the Sharpe ratio is stable to minor errors in those views Conclusion section.
Ruth: The authors note that while they showed this improvement, future work should focus on producing such views in practice Conclusion section.
Warren: So, to wrap up the paper "Directionally Correct Views in Portfolio Optimization," it seems the main point is that directionally correct expert information can help improve portfolio design within Markowitz models.
Iris: They also demonstrate that even when we move to more complex end-to-end models, simpler Markowitz-type models can perform just as well once those directionally correct experts are incorporated Abstract.
Ruth: The authors discuss how to incorporate expert views in end-to-end portfolio models using entropy based ideas Abstract.
Warren: So the implication here is that you don't always need the most complex model to capture the benefits of directional information if you have good views, and that this works even with simpler models.
Iris: And they show that the Sharpe ratio is stable to minor errors in those expert views Conclusion section.
Ruth: It’s about making sure we align the available historical data with conditions expected over the investment horizon Abstract.
Warren: That's what this paper is really doing, showing how to bridge that gap using directional information.
Conclusion: Ruth: So we're wrapping up this look at "Directionally Correct Views in Portfolio Optimization." The main idea is that when you use expert opinions, or views, you can actually improve how you build your investment portfolio using historical data.
Warren: Right, so it's about taking those expert ideas and seeing if they make a real difference in the actual math we use for optimization.
Iris: What the paper shows is that even a small amount of information that’s directionally correct can materially improve investment decisions in a simple Markowitz setting.
Ruth: It seems like they've established this framework using the Sharpe ratio, defining it as w top mu minus rf over square root of w top sigma w, and then incorporating expert predictions for the mean and covariance.
Warren: That formula is specific—it involves an angle theta mu,sigma(nu, omega), which basically tells you when that realized Sharpe ratio hits its maximum point.
Iris: Proposition one proves there's a unique angle where the inverse of omega equals some constant times the inverse of mu sigma for that maximum.
Ruth: And they show how mean views fit into the Markowitz framework by changing the historical measure, moving from P to Q, which is defined by a constrained problem.
Warren: So, if we assume returns are normal under P, they also become normal under Q with the new mean view and covariance.
Iris: Then Theorem two applies to this scenario and shows that the Sharpe ratio S lambda is nondecreasing on the interval zero to one when nu lambda equals nu plus lambda of mu minus nu.
Ruth: Plus, there's a condition for concavity around one, which is A plus three B greater than or equal to zero.
Warren: And then they move into this more complex end-to-end model where those views are incorporated through an exponential change of measure over a lookback window.
Iris: The network is trained by gradient ascent on the view-adjusted Sharpe ratio S phi lambda, which involves the term m lambda phi over p v lambda phi.
Ruth: The experiments with daily NSE India data from two thousand six to twenty twenty five show that both models improve sharply as lambda increases, with most of that gain happening at small lambdas.
Warren: It looks like the two Sharpe curves are strikingly similar throughout the whole process, which is pretty interesting for comparing these approaches.
Iris: The conclusion is that incorporating directionally correct expert views can improve portfolio design, and the Sharpe ratio stays stable even with minor errors in those views.
Ruth: So what this means for us right now is that you don't need the most complicated model to get a better portfolio if you have good directional information.
Warren: It suggests that aligning our historical data with future expectations through expert views is a viable path forward.
More episodes
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- The day in conversation for 2026-11-18
- Daily Summary for 2026-11-18
- 2601.05274-On the use of case estimate and transactional payment data in neural networks for individual loss reserving
- 2009.00868-Loss-Given-Default Modeling by Post-Last Passage Time Process
- 2606.30070-Financial Resilience Evaluation: From Conditional Expectations to Dynamic Convex Risk Measures
- 2212.12687-Measuring price impact and information content of trades in a time-varying setting
- Daily Summary for 2026-11-17
- 2401.08323-Dynamic portfolio selection under generalized disappointment aversion