Construct the choice experiment
Constructed seven sector profiles and validated 21 pairwise choices per respondent, with no missing responses in the four-person sample.
02 / Research notebook
What a choice reveals. What a model cannot.
Recovering investor risk preferences from JSE sector returns.
Collaborative BSc Honours research · University of the Free State- Degree
- BSc Honours, Risk Analysis
- Co-researcher
- Athini Mahlangeni
- Supervisor
- Dr. Morné Sjölander
- Status
- Submitted, September 2026
4Respondents
84Recorded choices
7JSE sectors
Research completed.
Dissertation submitted.
September 2026
01 / The question
Can preferences be
recovered from choices?
A model can describe risk. Understanding how someone responds to it is a different question.
A four-respondent study of investment choices across seven JSE sectors. Each respondent made 21 paired choices between sectors described by historical return and downside risk, giving 84 recorded choices. Respondent-specific choice models were fitted, validated out of sample and stress-tested for sensitivity to the sector set. The dissertation was submitted in September 2026.
02 / The process
Estimate individual preferences
Fitted separate choice models using alternative risk representations: volatility, 95% Value at Risk, Expected Shortfall and composite risk.
Challenge the estimates
Assessed separation, model conditioning, individual-pair influence and predictive performance, then tested how removing sectors changes the estimates.
03 / The evidence
The useful answer
includes the uncertainty.
Two respondents showed a consistently positive return preference across the fitted risk specifications. Risk-coefficient intervals included zero for all four respondents, and predictive performance varied across individuals.
These conclusions rest on four respondents and 21 comparisons each. They describe those individuals, not investor behaviour in general; more comparisons per respondent and a larger sample would be needed to pin down risk preferences. The research was a collaboration with Athini Mahlangeni, supervised by Dr. Morné Sjölander.
04 / The conclusion
Choices reveal
some of the answer.
Partly. Where a respondent’s choices were consistent, 21 paired comparisons were enough to recover an interpretable preference for return that also predicted held-out choices. Recovering attitudes to risk was a harder problem than this design could resolve.
- Return preferences were recoverable for two respondents. Their return coefficients were positive and statistically different from zero across every risk specification, and their models beat the 50/50 benchmark out of sample: 90.5% and 85.7% accuracy, with log losses of 0.256 and 0.384 against 0.693.
- The other two respondents’ choices did not produce a model that improved on chance, which is itself informative: not every set of choices carries a stable return–risk trade-off.
- Risk attitudes stayed uncertain. Three respondents had negative risk coefficients and one positive, but every risk interval included zero.
- Rankings were stable even when utility weights were not. Bradley–Terry and Thurstone–Mosteller placed Gold first for all four respondents, while leave-one-sector-out checks showed the attribute-based estimates depended heavily on which sectors were included.
- Very consistent choosers break standard estimation. One respondent’s strictly return-maximising choices caused complete separation, and ridge regularisation was needed to obtain finite, stable estimates.
Explore the research
An interactive view
of the research.
Choose between pairs of JSE sectors, described by historical return and downside risk, and the dashboard estimates the risk preferences your choices reveal, using the same choice model as the study.
Your choices stay in your browser. Nothing is stored or sent to a server.
05 / The methods
A closer look
under the surface.
- Paired-comparison design
- Conditional logit
- Ridge regularisation
- GARCH-family volatility models
- Extreme Value Theory
- Leave-one-pair-out validation
- Leave-one-sector-out sensitivity
- Bradley–Terry benchmarks