Without knowing too much about the context, I would say actual value. Because generally speaking, what you're trying to figure out is the error in the forecast.
E.g. if I'm forecasting temperature, and I say it's going to reach a daily high of 15 degrees Celsius tomorrow. The actual high was 13.5 degrees Celsius.
Percentage error of the forecast = (forecast - actual)/actual = ?
your error is forecast - actual = (15-13.5) = +1.5 (the forecast was 1.5 degrees too high)
The percent error is 100 * (1.5/13.5) = 11.111%
So, as a percentage of the true value, your forecast was 11.1% too high. This is the way of reporting things that makes the most sense to me.
If you did the opposite, you'd compute -1.5/15 = -10%. Your true value was 10% lower than the forcasted value. This has meaning too, but is not really an error in the forecast. It's treating reality as a thing that deviated from the forecast.