An automated search is carried out for microlensing events using a catalogue of 44554 variable superpixel lightcurves derived from our three-year monitoring program of M31 .
Each step of our candidate selection is objective and reproducible by a computer .
Our search is unrestricted , in the sense that it has no explicit timescale cut .
So , it must overcome the awkward problem of distinguishing long-timescale microlensing events from long-period stellar variables .

The basis of the selection algorithm is the fitting of the superpixel lightcurves to two different theoretical models , using variable star and blended microlensing templates .
Only if microlensing is preferred is an event retained as a possible candidate .
Further cuts are made with regard to ( i ) sampling , ( ii ) goodness of fit of the peak to a Paczyński curve , ( iii ) consistency of the microlensing hypothesis with the absence of a resolved source , ( iv ) achromaticity , ( v ) position in the colour-magnitude diagram and ( vi ) signal-to-noise ratio .
Our results are reported in terms of first-level candidates , which are the most trustworthy , and second-level candidates , which are possible microlensing but have lower signal-to-noise and are more questionable .
The pipeline leaves just 3 first-level candidates , all of which have very short full-width half-maximum timescale ( t _ { 1 / 2 } < 5 days ) and 3 second-level candidates , which have timescales t _ { 1 / 2 } = 31 , 36 and 51 days .
We also show 16 third-level lightcurves , as an illustration of the events that just fail the threshold for designation as microlensing candidates .
They are almost certainly mainly variable stars .

Two of the 3 first-level candidates correspond to known events ( PA 00-S3 and PA 00-S4 ) already reported by the POINT-AGAPE project .
The remaining first-level candidate is new .
This algorithm does not find short-timescale events that are contaminated with flux from nearby variable stars ( such as PA 99-N1 ) .