Volatility forecasting in practice: exploratory evidence from European hedge funds

Abstract

This note provides survey evidence of volatility forecasting practices in a number of European hedge funds. Results confirm the academic consensus that option-implied volatility (IV) is a commonly used risk management and volatility forecasting tool among “sophisticated” investors, but also highlight the great popularity of simple historical models, whereas stochastic models are of lesser relevance. Sensible, market sentiment capturing forecasting solutions that reduce model complexity are not only demanded, but are also already implemented by a number of practitioners. The development of multi-model forecasting solutions that combine historical and IV information into a reliable predictor of volatility appears to be a promising path for research.

Appendices

Appendix 1: Model use in different markets by model categories

This table shows response and case frequencies to the question: Irrespective of forecasting horizons which category of models (multiple selection possible) is most relevant to your fund in regard to the following markets?

Absolute numbers	IV	HISVOL	GARCH	SV	Total responses	Cases
Overall	58	58	25	17	158	106
Equity	45	45	18	12	120	81
FX	5	5	3	1	14	8
Commodities	3	3	1	2	9	6
Debt	2	4	1	1	8	6
Derivatives	3	1	2	1	7	5

Relative numbers	IV (%)	HISVOL (%)	GARCH (%)	SV (%)	Total responses (%)	Cases (%)
Overall	55	55	24	16	149	100
Equity	56	56	22	15	148	100
FX	63	63	38	13	175	100
Commodities	50	50	17	33	150	100
Debt	33	67	17	17	133	100
Derivatives	60	20	40	20	140	100

Appendix 2: Model use over different forecasting horizons in equity markets

This table shows response and case frequencies to the question: In regard to equity markets which category of models (multiple selection possible) is most relevant to your fund to forecast volatility up to a day, week, month, year and beyond one year?

Absolute	IV	HISVOL	GARCH	SV	Not relevant	Total responses	Cases
1 day	7	5	3	2	20	37	32
1 week	12	8	5	2	15	42	33
1 month	13	12	6	3	14	48	36
1 year	8	11	2	2	17	40	33
> 1 year	5	9	2	3	18	37	31
Total	45	45	18	12	84	204	165

Relative	IV (%)	HISVOL (%)	GARCH (%)	SV (%)	Not relevant (%)	Total responses (%)	Cases (%)
1 day	22	16	9	6	63	116	100
1 week	36	24	15	6	45	127	100
1 month	36	33	17	8	39	133	100
1 year	24	33	6	6	52	121	100
> 1 year	16	29	6	10	58	119	100
Total	27	27	11	7	51	124	100

Appendix 3: Forecasting approaches over different forecasting horizons in equity markets

This table shows recoded survey response from Appendix 2. If a respondent indicated that s/he regards HISVOL and IV as most relevant to predict volatility over a day, then this answer is recoded into the single-variable HISVOL-IV (and so forth). This is done over all forecasting horizons, before the sum over all horizons is taken as an overall indication for model combinations in equity markets; thus, response numbers equal cases analysed.

Absolute	Day	Week	Month	Year	> Year	Overall
Single-Model Approach	9	13	13	10	9	54
IV-Single	5	8	8	4	3	28
HISVOL-Single	3	3	3	5	5	19
GARCH/SV-Single	1	2	2	1	1	7
Multi-Model Approach	3	5	9	6	4	27
HISVOL + IV	0	1	3	4	1	9
HISVOL + GARCH/SV	1	1	4	2	1	9
IV + GARCH/SV	1	0	0	0	1	2
HISVOL + IV + GARCH/SV	1	3	2	0	1	7
Total	12	18	22	16	13	81

Relative	Day (%)	Week (%)	Month (%)	Year (%)	> Year (%)	Overall (%)
Single-Model Approach	75	72	59	63	69	67
IV-Single	56	62	62	40	33	52
HISVOL-Single	33	23	23	50	56	35
GARCH/SV-Single	11	15	15	10	11	13
Multi-Model Approach	25	28	41	38	31	33
HISVOL + IV	0	20	33	67	25	33
HISVOL + GARCH/SV	33	20	44	33	25	33
IV + GARCH/SV	33	0	0	0	25	7
HISVOL + IV + GARCH/SV	33	60	22	0	25	26
Total	100	100	100	100	100	100

Appendix 4: Qualitative factors influencing model choice

After participants ranked model categories according to perceived forecasting performance, we posed the following open-end question: Why do you think is your highest ranked model superior to its competitors? The table below reports survey responses (as given and not amended) along with our coding of participants’ answers.

Highest ranked	Respondent		Coding
HISVOL	HISVOL.1	There is an underlying mean-reversion supported by implicit valuation	Model Characteristics
HISVOL	HISVOL.2	Simplicity of use	Simplicity
HISVOL	HISVOL.3	For my purposes “sensible” is more important than “accurate” and it is important to avoid the excess fears that are built into IV after big market events so I prefer a simple historic estimate. I’m not convinced Garch or SV add value over multi-month periods	Model Characteristics Simplicity Limitations of other models
HISVOL	HISVOL.4	Simple and less sensible to overfitting	Simplicity Model Characteristics
HISVOL	HISVOL.5	Because, as stupid as it is, it is what is used by the majority of market-makers. GARCH is used by risk managers, and SV is used only in vanilla-exotic derivatives pricing (but is arbitrageable because it places the price quote off-market). IV is driven by Keynesian ‘animal spirits’ in equity markets, and is a poor predictor of ex post realised volatility, as well as consistently arbitrageable (selling rich straddles vs. buying cheap wings). What matters in the market is NOT what one thinks; what matters is what the preponderant majority of other participants think—and they tend to think along HISVOL lines. I have made more than 50,000 multivariate and exotic option prices during my 20-year career, and I was able to be on-market using HISVOL as opposed to anything else (including ‘fits’ like Dupire local vol, which contain no scientific drivers)	Industry Use Limitations of other models
HISVOL	HISVOL.6	I disbelieve in general in statistical prediction, so the Historical vol analysis should be less deceiving	Limitations of other models
IV	IV.1	Historically it has given superior forward predictions	Predictive Ability
IV	IV.2	It is capturing market participants positioning	Capturing Market Sentiment
IV	IV.3	When adjusted for risk premium IV captures market participants views in aggregate	Capturing Market Sentiment
IV	IV.4	I think it’s SV.	n/a
IV	IV.5	The market is always right, especially in markets where volatility is actively traded. The problem with volatility models is that they have to be consistent both on the level of the underlying and the level of volatility. When both are traded, they usually cease to make sense. In addition, complex models do not make a complex phenomenon easier to understand, they just hide the many additional assumptions required	Capturing Market Sentiment Limitations of other models
IV	IV.6	Forward looking	Predictive Ability
GARCH	GARCH/SV.1	Practical use in the field	Industry Use
GARCH	GARCH/SV.2	Works better out of sample	Model Characteristics
SV	GARCH/SV.3	The classification is little bit too simple for how I view vol models in general. Hisvol will under- or over-estimate vol and underperform due to equal weighting schemes, so ranked it worst. But I find the $ value or SR contribution from GARCH or SV little at best in practice, when compared with easier EWMA-type models. Not sure which one contains more info IV or backward looking models, that is probably debatable, but combining IV and EWMA(or GARCH/SV models) tends to work better (backward and forward combination). Also I prefer range volatility to squared return models (even for RQV type models at high frequency). Finally, all the models above fail forecasting vol horizons beyond a week or so at best, because they tend to capture tail shocks mostly, but they are relatively silent for cyclical or slow-moving component of vol where excess kurtosis is not so apparent. For long-term forecasting vol, I tend to use two-factor vol models (slow moving factor + tail shock factor) and also use fundamental variables n the equations	Limitations of other models
Unranked	Unranked.1	We don’t use models we trade markets. People who trade models normally end up working for banks and imploding them. No I’m being serious