We use full valuation of derivatives portfolios using scenarios from historical data.
For simple contracts, this is relatively fast.
For contracts requiring monte carlo simulation, this becomes painfully slow for large portfolios.
I am thinking about ways of improving the current situation. Most likely, in my mind, this would involve a full valuation done over night, resulting in PnL vectors per contract and possibly other results.
During intraday, as market data changes, these "base PnLs" could probably be used along with any other previously calculated result to get a decent approximation without having to perform a full valuation.
I am looking for some existing work making these ideas concrete (or possibly suggesting a different approach entirely...).
A basic idea would be to compute sensitivities alongside the full valuation. In this way, the PnLs will be accompanied by, say, delta and gamma sensitivities, and these could be used in combination with the new market data to get an updated PnL vector.
The question is if this is feasible or indeed if it's any better than simply using a delta-gamma approximation to begin with?
Another idea could be to rely on sensitivities implicitly by simply valuing the contract under a discrete set of price points with the current price in the middle, say [-3%, -2%, -1%, -0.5%, 0, ...], and performing interpolation for finding a fair value intraday.
Any feedback is welcome.