If you don't have information of the phase of the signals, you'll not be able to reconstruct the waveform.

I don't know what your "Excel and Fourier analysis tool" is, but it should give you frequency and magnitude as well as phase information. Otherwise you're lost...

If you have done Fourier over time sample sets, when doing the reverse with the same step size, you will get the same output in theory, such is used in audio compression. However, I don't think this is what Simon is asking about. I believe he has a fixed single set of samples, with a single Fourier analysis of that 1 complete chunk. Yes, the reverse can be closely approximated, depending on the sample size and depth of the Fourier analysis, though, there will be loss.

Now, onto Simon's gut of the question. The most accurate point in the Fourier should be the center of your sample set. IE, if you do an analysis of 1000 samples. When reversing the Fourier, the center at sample 500 should be the most accurate, where as you go further before and after sample 500, the error in reproduction will grow. This is assuming you are using a true Fourier conversion of the entire sample set from beginning to end. This will not be true for some sequential Fourier conversion algorithms designed/used in sound or streamed sampled signals.

To answer Simon's question, does Excel's Fourier analysis tool equally weigh all the samples in his sample set. If yes, then, using an equivalent reverse tool, the reconstructed samples will be 'in-phase' or, most accurate from the center sample, progressively loosing refinement as you spread out to the left and right of center. Now, if your sample set is, say, only 200 samples or less, with integer values from -100 to +100, every value may be perfectly reconstructed with a well enough defined Fourier analysis and reconstruction. Doing this with 10000 samples, with random integers from -10000 to +10000, you will not be able to get the true data back unless you have an absurd size of points & definition in your Fourier, but it is doubtful. But you will get an approximation.

Such tools are used to predict future trends in many applications, such as the stock market, and those new bloody HFT (search 'high frequency trading') machines equipped with tons of NVIDIA commercial graphics cards doing nothing but FFT an top of FFT again and again with real-time stock data to gain an edge on the stock market. It should be illegal.