Welcome to the experimental world!
To be fair, it's quite normal that oscillations do not perfectly coincides after some periods. You have a period of the oscillation $T \simeq 0.7 \, s$ and a shift (if minimal) of $0.3 \, s$ after $300 \, s$. This means an approximate error on the period of $\Delta T = 0.3 \, s \cdot \frac{0.7 \, s}{300 \, s} = 7 \cdot 10^{-4} \, s$, i.e. a relative error of $\frac{\Delta T}{T} = 1 \cdot 10^{-3} = 0.001$, i.e. a $.1 \%$ error.
Structural dynamics usually deal with modes (shapes, frequencies and damping coefficients) of a structure, and not with time histories over long time. I could say that $.1\%$ relative difference between the natural frequency of the numerical model and the experimental model is good enough for everyone.
Quantify your error.
Example 1 - Undamped oscillator. For an undamped oscillator with mass $m$ and stiffness $k$, the natural angular frequency and the period are
$$\omega = \sqrt{\frac{k}{m}} \quad , \quad T = \frac{2 \pi}{\omega} = 2\pi \sqrt{\frac{m}{k}} . $$
Let's assume here for simplicity that the only source of errors are different values in $m$ and $k$, and no other factors not included in the model. The error on $\omega$ due to a small difference in the value of $m$ and $k$, with the sensitivity analysis, reads
$$
\Delta \omega = \frac{\partial \omega}{\partial k} \Delta k + \frac{\partial \omega}{\partial m} \Delta m
= \frac{1}{2}\frac{1}{\sqrt{km}} \Delta k - \frac{1}{2} \sqrt{\frac{k}{m^3}} \Delta m $$
$$
\Delta T = \frac{\partial T}{\partial k} \Delta k + \frac{\partial T}{\partial m} \Delta m
= -\pi\sqrt{\frac{m}{k^3}} \Delta k + \pi \sqrt{\frac{1}{km}} \Delta m $$
while the relative error is obtained, dividing by $\omega$ and $T$ respectively,
$$\frac{\Delta \omega}{\omega} = \frac{1}{2} \dfrac{\Delta k}{k} - \frac{1}{2} \frac{\Delta m}{m} \ .$$
$$\frac{\Delta T}{T} = -\frac{1}{2} \dfrac{\Delta k}{k} + \frac{1}{2} \frac{\Delta m}{m} \ .$$
Thus, with perfect mass, a $+1\%$ relative error on the measure of the stiffness produces a $+0.5\%$ relative error on the angular frequency; with a perfect stiffness, a $1\%$ relative error on the mass produces a $-0.5\%$ relative error on the angular frequency.
The $.1 \%$ relative error you get on the period, it would be produced by a $.2 \%$ relative error on the mass (on the stiffness) for perfect stiffness (mass).
Example 2 - Damped oscillator. You could do the same analysis for a damped oscillator, including the damping coefficient in the analysis.
Damping is usually harder to be modeled since it comes from different processes that are only included with a damping coefficient (usually a too simplistic model).
Possible sources of error
It's likely that:
- some parameters of your system don't "precisely" coincide with the values you used in the analytical model/formula;
- some features of your system are not "perfectly" represented by your model: this is a typical case for:
- physical constraints and boundary conditions of models in structural dynamics problems, like the one of your interest;
- damping from air, or structural damping
The meaning of world "precisely"
What do we mean by "precise"? We usually mean "precise enough" for the application/the tolerance we need. It's likely that you can't build a numerical model that gives the very same results of experimental tests, given all the possible sources of errors, and uncontrollable uncertainty. When dealing with critical applications, some safety coefficients are used.
Approach to experimental tests and tuning of numerical models
A numerical model of a physical system is usually naturally affected by modeling errors. It's a good practice to have some "tunable" parameters in the numerical model and tune them (parameter identification) to make the prediction of the numerical model match with the measurements performed during some experimental test.