CONCLUSION

I have presented here ways in which two classical spectral-analysis adaptive algorithms can be modified so that they are suitable to any prediction problem. The first one, the LSL algorithm, solves an exact least-squares problem on windows of increasing size; its residuals don't depend on future data. The other one, a Burg-type algorithm, uses a tapering of the residuals to make the process adaptive, but seems less rigorous mathematically. However, the better results are obtained by the Burg-type algorithm, which is better stabilized than the LSL algorithm.