Dummy variable confusion, need panel help

[Max25]

Hi,
I am getting a strange regression result with a dummy variable which I assume is not possible and now I am doubting whether what I am doing is right.
Here is a brief overview of what I am doing.
I analyze fund price deviations from net asset value and want to see what explains these deviations. Sometimes mutual funds trade at higher or lower prices than what their underlying assets are worth (for nonfinance people )
i use daily data over 10 funds, create a panel, and apply some variables including a dummy that is 1 over the recent financial crisis period.
my dependent variable is the log difference between fund price and the net asset value. if the resulting value is positive its a premium, if its negative a discount. because i only care about what explains the deviation from net asset value i use the absolute difference between the two

the image below shows the development of premiums and discounts (not absolute values of course) over the period... the dummy is set to 1 over the 2007-08 period when the deviations widen


now i wonder how it is possible that in the final regression i get a negative coefficient for my crisis dummy that looks like this:
Coefficient Std. Error t-Statistic Prob.
DUMMY -0.001544 0.000271 -5.704378 0.0000

wouldn't this imply that the existence of the crisis decreased the deviation and brought the price closer to the fundamental value?
if so, what could i have done wrong with my model? from the looks of that graph the crisis period dummy should give me a positive coefficient or shouldn't it?

since I am no expert please find the model specification below

I used a fixed-effects model
with cross-section GLS weights and coef covariance method set to white cross-section
however i should note i am only partially sure that the latter two are actually correct, however even when i use cross-section SUR on the latter two i still get a negative coefficient

below the header of my eviews output:
Method: Panel EGLS (Cross-section weights)
Periods included: 1266
Cross-sections included: 20
Total panel (balanced) observations: 22002
Linear estimation after one-step weighting matrix
White cross-section standard errors & covariance (d.f. corrected)

i am really desperate and have no idea what to think. if you have any comments, even if its just saying that my model estimation isn't completely useless, that would be greatly appreciated!
thanks in advance

[RobHayward]

I don't think this should be a surprise. I think that the dummy variable is telling you that the fund discount to asset value increased in the crisis. This would make sense if there is panic selling of funds while equity valuations are supported by physical assets or more liquid markets. The diagram that you post seems to show increased volatility of the gap between asset value and price of fund, with (I think) more negative than positive readings.

[Max25]

but i am looking at absolute deviations from net asset value in my regression. shouldn't the negative coefficient for the dummy then suggest that the deviation from the fundamental value was actually decreased by the size of the coefficient (-0.001544) and hence making it more efficiently priced?

i greatly appreciate any clarification

[RobHayward]

Sorry - missed the 'absolute'. It seems strange, unless there was a premium before and a smaller discount afterwards.

[Dataminer]

i use daily data over 10 funds, create a panel, and apply some variables including a dummy that is 1 over the recent financial crisis period.

Max25
Just a quick reaction based on graph and quote. Is it the case that the dummy is an expression of a different mean in the series but you are interested in a different variance i.e. you need some sort of heteroskedasticity specification.
Gerald

[Max25]

Max25
Just a quick reaction based on graph and quote. Is it the case that the dummy is an expression of a different mean in the series but you are interested in a different variance i.e. you need some sort of heteroskedasticity specification.
Gerald

i would appreciate some further clarification on what you think i should do

[Max25]

any other comments? i really appreciate the help

[Dataminer]

Max25
Just a little more thought on my previous contribution. I guess the graph is just one member of the panel and you propose to model all members demonstrating similar behaviour Suggest the following - no detailed experience to guage suitability of these.
(i) Analyse each series as a (G)ARCH. this would show if the changing variance can be attributed to a stationary AR process or not.
(ii) Analyses each series looking for an optimal structual break by varying your dummy variable date definition but use it as a sort of time slice segmentation in a Goldfeld-Quandt test for heteroskedasticity to find the date of the most significant segmentation. This will require a goodly number of observations in each calendar segment.
(iii) try to apply the above in a panel setting.

Gerald