Monte Carlo: generating autocorrelated data from empirical distribution my problem is the following: having a distribution function of daily casfhlows resulting from electricity trading, I need to calculate a yearly 99% VaR, i.e. the 1% percentile of yearly casfhlows distribution. 
Method 1/ Use CLT - fast and easy, albeit daily cashflows show some autocorrelation (0.3, 0.2 on the first two lags), hence assumptions are not satisfied.
Method 2/ Simulate as though daily cashflows are independent, i.e. 10 000 times generate 365 random values from the distribution function of daily cashflows and sum them. Like this, I get 10 000 yearly cashflows and get the respective percentile. 
Method 3/ Simulate taking account autocorrelation of daily cashflows. This is what I don't know how to do. Having an empirical distribution function X(avg, sdev) how can I successively generate random realizations while ensuring that the generated data have (at least on the first lags) the same autocorrelation structure as the empirical daily cashflows that I have? Is there a way of doing this without estimating an ARMA model? The reason I want to avoid ARMA is that once I randomly generate one observation and model the following observations via the estimated ARMA process, the distribution of predicted cashflows will drift away from the empirical distribution of daily cashflows (due to the expected bad fit of the ARMA process). 
Any hint is highly appreciated, 
Daniel
 A: For the sake of simplicity, consider that you invest in a portfolio of futures contracts and want to compute the VaR. You would first model the path of the futures prices and then calculate the distribution of the profits of the portfolio at the horizon given some arbitrary holdings. 
The second part is straight-forward, but the first part is generally more complicated with more options. The primary inputs into the futures curve would be the spot price and rates that drive the term structure of the futures curve (so the risk-free rate plus the convenience yield). You can interpolate those rates to arbitrary periods that are constant through time and then focus on modeling these. 
So to model the prices and these interpolated rates, I would focus on the log differences for the prices and the differences in the yields (these would be time-homogenous invariants). One obvious option is to estimate a VAR with a number of lags large enough to account for the seasonality in the spot price. For a daily series, that might be a lot of lags, so weekly or monthly might be better. Another alternative is to do a bootstrap that selects more than one period historically. For instance, if you can identify when the seasons are for your electricity data, then you can sample the whole season historically. Alternately, you can sample each month only from previous months (so January's bootstrap is always from a January historically).
After simulating you would have to build back up to the futures prices at your horizon and then the distribution of the portfolio at the horizon. Then you would calculate VaR from this distribution. 
