[R-sig-finance] Is oanda.com data trustworthy?

[Ajay Narottam Shah]

> But as you know, FX markets operate 24hrs, and are broker-based implying that
> there will be a) different quotes throughout the day and no "final" one and
> b) different sources may well disagree about quotes at one point [ though
> arguably not by the margin you show there ]. Oanda may report London fixings
> whereas the US Fed probably calls up US banks and end-of-day US time.

I was aware of the 24hour issue. I thought it may lead to some clear
leads and lags. Read on.

> As for the question above, maybe time to test some Granger causality between
> Oanda and the FED sources. :-)

> usfed <- mydb("currency", "EUR/USD", firstdate="2002-01-01",
                lastdate="2004-12-31")
> oanda <- window(read.zoo("EUR", skip=1), start=as.Date("2002-01-01"),
                end=as.Date("2004-12-31"))
> prices <- cbind(usfed, oanda, all=FALSE)
> plot.zoo(prices, plot.type="single", col=c("blue", "yellow"), lwd=2)

# Visually looks okay.

> returns <- prices2returns(prices)
> 
> summary(returns)
     Index                usfed              oanda         
 Min.   :2002-01-03   Min.   :-1.82563   Min.   :-1.80702  
 1st Qu.:2002-09-26   1st Qu.:-0.46976   1st Qu.:-0.42765  
 Median :2003-07-01   Median :-0.04195   Median :-0.05355  
 Mean   :2003-07-02   Mean   :-0.03983   Mean   :-0.05427  
 3rd Qu.:2004-04-06   3rd Qu.: 0.31539   3rd Qu.: 0.30200  
 Max.   :2004-12-31   Max.   : 2.11408   Max.   : 1.98718  

# Looks plausible.

> returns <- as.data.frame(returns)
> 
> plot(returns$oanda, returns$usfed)

It looks like a zero correlation!!!!!!!!!!!!

> summary(lm(usfed ~ oanda, returns))

Call:
lm(formula = usfed ~ oanda, data = returns)

Residuals:
     Min       1Q   Median       3Q      Max 
-1.78081 -0.42691  0.01766  0.36543  2.10784 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.03637    0.02563  -1.419    0.156
oanda        0.06373    0.04205   1.516    0.130

Residual standard error: 0.6191 on 586 degrees of freedom
Multiple R-Squared: 0.003904,	Adjusted R-squared: 0.002204 
F-statistic: 2.297 on 1 and 586 DF,  p-value: 0.1302 

> cor(returns)
           usfed      oanda
usfed 1.00000000 0.06248328
oanda 0.06248328 1.00000000

This is truly scary - a returns correlation of 0.06 between two
measures of returns!

> library(lmtest)
> grangertest(returns$usfed ~ returns$oanda, order=5)
Granger causality test

Model 1: returns$usfed ~ Lags(returns$usfed, 1:5) + Lags(returns$oanda, 1:5)
Model 2: returns$usfed ~ Lags(returns$usfed, 1:5)
  Res.Df  Df      F Pr(>F)
1    572                  
2    577   5 1.2864 0.2681
> grangertest(returns$oanda ~ returns$usfed, order=5)
Granger causality test

Model 1: returns$oanda ~ Lags(returns$oanda, 1:5) + Lags(returns$usfed, 1:5)
Model 2: returns$oanda ~ Lags(returns$oanda, 1:5)
  Res.Df  Df      F    Pr(>F)    
1    572                         
2    577   5 59.686 < 2.2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 

So there IS something going on! Could be a time of day story?

Here are a few days of returns, to help fix the intuition:

> tail(returns, 10)
                 usfed       oanda
2004-12-16  1.06300992 -0.76106917
2004-12-17 -0.06020826  1.18581098
2004-12-21  0.14203074 -0.42746525
2004-12-22 -0.09716418  0.12040941
2004-12-23 -0.75145123 -0.12040941
2004-12-24 -0.33297193 -0.91447652
2004-12-28  0.02928946 -0.58364604
2004-12-29  0.38988948  0.05443658
2004-12-30 -0.40446502  0.04080800
2004-12-31  0.62592084 -0.29964610

So there does seem to be a pattern where information for 2004-12-16
with the US Fed (1.06%) seems to show up on the next day for oanda
(1.18%).

If I lag the US Fed data by 1, the correlation goes up to 0.57. I
think this is a useful discovery :-) but .57 still sux!

-- 
Ajay Shah
ajayshah at mayin.org
http://www.mayin.org/ajayshah