📈 economics

Joint Identification of Beliefs and Preferences from Point and Density Forecasts

This paper develops a decision-theoretic framework demonstrating that the joint observation of point and density forecasts in survey data enables the separate identification of agents' beliefs and preferences, revealing richer information about expectation formation than previously recognized.

Original authors: Ayush Jha, Frank Fabozzi

Published 2026-07-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ayush Jha, Frank Fabozzi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Crystal Ball and the Crystal Ball's Shadow

Imagine you are trying to guess what the weather will be like next week. You could just say, "It will be 75 degrees." That's a point forecast: a single, specific number. Or, you could say, "There's a 20% chance of rain, a 50% chance of sun, and a 30% chance of clouds." That's a density forecast: a full picture of all the possibilities and how likely they are. For decades, scientists who study how people think (economists and psychologists) have looked at these two types of guesses separately. They've asked, "Are the single numbers accurate?" and "Are the probability pictures well-calibrated?" But they rarely asked what happens when you look at both at the same time.

This paper dives into that missing link. It treats these forecasts not just as guesses, but as clues left behind by a person's mind. The authors are interested in two hidden things: Beliefs (what you think will actually happen) and Preferences (how you feel about those outcomes—do you love risk, or do you hate it?). Think of beliefs as the map of the territory, and preferences as the compass that tells you which direction you want to go. Usually, when you see someone make a choice, it's hard to tell if they chose a path because they thought it was the only one available (belief) or because they really wanted to go there (preference). This paper asks: If we see both the map (the density forecast) and the final destination they picked (the point forecast), can we finally separate the map from the compass?

The Detective Work: Unmasking the Mind

The authors, Ayush Jha and Frank Fabozzi, set up a detective story where the "suspects" are the hidden beliefs and preferences of professional forecasters. They use a framework called Rank-Dependent Utility, which is a fancy way of saying: "People don't just look at the odds; they twist them in their heads before making a decision." Maybe you think a 10% chance of winning is actually more exciting (or scarier) than it mathematically should be.

Here is the big twist the paper reveals: You can't solve the mystery with just one type of clue.

If you only have the density forecast (the full probability map), you know exactly what the person believes will happen. You know they think there's a 30% chance of rain. But you have absolutely no idea if they are a risk-taker who loves the rain or a risk-avoider who hates it. The map is clear, but the compass is invisible.

If you only have the point forecast (the single number, like "It will be 75 degrees"), you are stuck in the dark. That single number could be the result of a person who believes it will rain but doesn't care, or someone who believes it will be sunny but is terrified of the rain. Without the map, you can't tell the difference between their beliefs and their feelings.

The Breakthrough: When you have both the map and the single number together, you get a powerful new tool. The map tells you the beliefs. The single number then acts as a filter, narrowing down the list of possible "compasses" (preferences) that could have led to that specific guess. It's like seeing a person's footprints (the map) and then seeing exactly where they stopped to rest (the point). You can now deduce how tired they were or how much they enjoyed the hike.

The Reality Check: Not All Clues Fit the Story

The authors don't just say, "We solved it!" They take a very careful, scientific approach. They test their theory using real data from the Survey of Professional Forecasters (SPF), a long-running survey where experts guess things like GDP growth and inflation.

Here is what they found, and it's a bit surprising:

Sometimes the story fits perfectly: For some variables, like the GDP deflator inflation, the experts' point forecasts and their density forecasts fit together beautifully. The math works out, and the "compass" (their preference for risk) looks normal and stable. The model can explain their behavior.
Sometimes the story breaks: For Real GDP growth, the clues don't fit. When the authors tried to force the math to work, the "compass" they had to invent was broken. It suggested that the experts had "negative risk aversion"—a fancy way of saying they were acting like they loved chaos and disaster in a way that doesn't make sense for a normal human being.

The paper argues that this isn't because the experts are weird. Instead, it suggests that the model itself might be missing something. The gap between what the experts said they believed (the density) and what they actually predicted (the point) is too wide for the current theory to explain. It's like trying to fit a square peg into a round hole; the peg isn't the problem, the hole is the wrong shape.

The Takeaway

The main lesson isn't that we can now perfectly read people's minds. The paper is careful to say that we can't separate beliefs and preferences without making some extra assumptions. Instead, the paper shows that forecast surveys are a treasure trove of information we've been ignoring.

By looking at the relationship between the "what if" (density) and the "what will be" (point), we can do more than just check if the forecast was right. We can check if the story the forecaster is telling makes sense. If the story doesn't add up (like with the GDP growth data), it tells us that our current understanding of how people make decisions under uncertainty might need an upgrade.

In short, the paper teaches us that a single number and a probability chart are best friends. Alone, they tell a partial story. Together, they reveal the complex, sometimes messy, and always fascinating way humans navigate an uncertain future. And sometimes, when they don't fit together, it's the universe's way of telling us that we need a better map.

Technical Summary: Joint Identification of Beliefs and Preferences from Point and Density Forecasts

1. Problem Statement

The paper addresses a fundamental identification problem in the analysis of expectation surveys. While point forecasts (e.g., expected inflation rates) and density forecasts (probability distributions over outcomes) are widely collected and analyzed separately, their joint observation has not been rigorously studied to determine what can be inferred about the underlying beliefs and preferences of forecasters.

Existing literature typically treats discrepancies between point forecasts and the central tendencies (mean, median, mode) of density forecasts as reporting errors or biases. This paper argues that such discrepancies may instead reflect the interaction between a forecaster's subjective beliefs and their decision-theoretic preferences. The core question is: Can beliefs and preferences be separately identified from the joint observation of a point forecast and a density forecast, and under what conditions?

The authors highlight that choices under uncertainty generally reflect both beliefs and preferences, making separation difficult without additional structure. While density forecasts appear to reveal beliefs directly, point forecasts are functional summaries of those beliefs that depend on the forecaster's objective (loss function or utility). Without joint analysis, neither object alone permits the separate identification of beliefs and preferences.

2. Methodology and Theoretical Framework

The paper develops a decision-theoretic framework to characterize the identification problem.

Formal Setup:

Observables: An agent reports a point forecast $p \in X$ and a density forecast $G \in \Delta(X)$ (a cumulative distribution function).
Latent Objects: The agent possesses a subjective belief $G$ and a preference structure defined by a utility function $u$ and a probability-weighting function $w$ .
Reporting Rule: The point forecast is generated by a mapping $T(G; u, w)$ , interpreted as a functional of the distribution (e.g., a certainty equivalent).

Identification Analysis:
The authors employ the language of partial identification (Manski, 2003; Tamer, 2010) and revealed preference rationalizability (Afriat, 1967; Varian, 1982).

Density Forecasts Alone: A reported density $G$ identifies the subjective belief exactly (under standard interpretation) but provides no information about preferences. The set of admissible preferences is the entire space $\mathbb{P}$ .
Point Forecasts Alone: A point forecast $p$ identifies neither beliefs nor preferences, as multiple belief-preference pairs can generate the same scalar summary.
Joint Observation: The joint observation $(p, G)$ restricts the set of admissible belief-preference pairs. The density forecast fixes the belief, and the point forecast imposes a constraint on the preference parameters consistent with that belief. However, the paper demonstrates that a single pair $(p, G)$ is generally insufficient for point identification of preferences because the preference space is infinite-dimensional, and the point forecast provides only a finite number of restrictions.

Rank-Dependent Utility (RDU) Framework:
To illustrate the results, the authors adopt the Rank-Dependent Utility model (Quiggin, 1982; Yaari, 1987). In this framework:

The density forecast determines the subjective distribution $G$ .
The probability-weighting function $w$ distorts $G$ into decision weights.
The point forecast is interpreted as a certainty equivalent ( $p = u^{-1}(V)$ ), where $V$ is the RDU value.
The authors derive conditions under which preferences can be identified from a panel of reports (varying beliefs over time) and establish order conditions (number of forecast rounds $\geq$ dimension of preference parameters).

Rationalizability vs. Identification:
A crucial distinction is made between identification (recovering parameters under a maintained model) and rationalizability (determining if any admissible parameter values exist that can explain the data). The paper proves that identification does not imply rationalizability; a model may impose identifying restrictions that result in parameter estimates violating standard admissibility conditions (e.g., negative risk aversion).

3. Key Contributions

Shift from Description to Identification: Unlike prior work that descriptively compares point forecasts to density statistics (e.g., checking if points equal medians), this paper formalizes the identification problem of recovering latent beliefs and preferences from forecast pairs.
Informational Complementarity: The paper establishes that point and density forecasts are informational complements. The density forecast disciplines beliefs, while the point forecast disciplines preferences. Joint observation strictly reduces the identified set of models compared to observing either in isolation.
Distinction Between Identification and Rationalizability: The authors formally distinguish between the ability to estimate parameters and the ability of a model to rationalize observed data. They show that "implausible" parameter estimates (e.g., negative values) may indicate a failure of the model to rationalize the data rather than evidence of unusual preferences.
Conditions for Identification: The paper provides necessary and sufficient conditions for point identification, including the requirement for a rich class of beliefs (e.g., a panel of forecasts with varying distributions) and parametric restrictions on the utility and weighting functions.

4. Empirical Results

The framework is applied to data from the Survey of Professional Forecasters (SPF), which reports both point and density forecasts for U.S. macroeconomic variables.

Methodology: The authors recover subjective distributions from SPF density forecasts using generalized beta densities. They estimate probability-weighting functions and then solve for the preference parameters (specifically the coefficient of relative risk aversion, $\gamma$ , under CRRA utility) that rationalize the observed point forecasts as certainty equivalents.
Findings:
- GDP Deflator Inflation: Implied preference parameters are positive, economically meaningful, and stable across different benchmark specifications. This suggests the joint forecasts are rationalizable within the RDU framework.
- Real GDP Growth: Under three of four benchmark specifications, the implied preference parameters are negative and statistically significant. According to the paper's Theorem 5, negative $\gamma$ violates admissibility conditions for CRRA utility. This indicates a failure of rationalizability: the observed point and density forecasts for real GDP growth cannot be jointly reconciled within the maintained RDU model.
- Core CPI and Unemployment: Results are mixed and less definitive, often yielding imprecise estimates or parameters that vary significantly with benchmark choices.
Interpretation: The negative estimates for Real GDP growth are not interpreted as evidence of risk-seeking behavior. Instead, they signal that the RDU model fails to capture the relationship between the reported beliefs and the point summary for this variable. The discrepancy likely stems from the density forecasts implying distorted means that differ materially from the reported point forecasts.

5. Significance and Claims

The paper claims that forecast surveys contain richer information about expectation formation than is typically recognized. By jointly analyzing point and density forecasts, researchers can:

Evaluate the internal consistency of forecast reports.
Test competing models of belief formation and decision-making under uncertainty.
Diagnose model misspecification or reporting conventions when standard decision-theoretic models fail to rationalize observed data.

The authors emphasize that their contribution is methodological rather than empirical. They do not claim to have estimated structural preference parameters as definitive truths. Instead, they demonstrate that the joint observation of forecasts imposes restrictions that can be used to assess the validity of decision-theoretic models. The empirical application serves to illustrate that while some forecast pairs are consistent with standard models, others imply a breakdown in the model's ability to explain the data, highlighting the need for alternative frameworks (e.g., ambiguity aversion) in future research.