Preparing CSV files for use in ArcGIS Desktop?
My quick fix is to create the first row all with dummy values, and then delete this row/record after bringing into in ArcGIS.
This first row contains representative values or often wildly different values (e.g. alphabetic characters even if the column contains numbers that I want to be text data type) and with the largest number of characters needed for that row (because text fields tend to get truncated).
Date/time values are subject to import errors (especially between Canada/U.S. default date formats) so my work around is to split the date/time parts in to separate columns (e.g. year, month, day, hour, minute), and then concatenate these in a new field calculation after successfully bringing into in ArcGIS.
The geographic coordinates tip from Jamie is also necessary - specify negative values for western hemisphere longitude and southern hemisphere latitude. And unicode takes care of special characters.
Lastly, if a field data type is still misinterpreted after bringing into ArcGIS I will add a new field in the correct data type and calculate/convert the values from the original field, but usually the dummy row/record takes care of most, if not all, problems.
A slight departure from the traditional CSV->ArcGIS layer could be to use ogr2ogr to create your shapefile from CSV BEFORE loading it into ArcMap.
When using OGR, you can manually create a CSVT file that describes your column types, similar to what ArcGIS attempts to do with the schema.ini file.
This post by @underdark on How to Specify Data Types of CSV Columns (for use in QGIS) explains the details of creating a CSVT. You can ignore the use of QGIS, as it just uses ogr to import the CSV data into it's UI.
With ArcGIS 10 you can create the feature class and set all field types first, then load the .csv file using ArcCatalog. Simply right-click the feature class and choose Load > Load Data. Then follow the Simple Data Loader dialog to match fields, etc. It is a simple loader in that there are not many options but it works for well-formatted data. I'm not sure if this is new to 10 because I don't have an earlier version handy to check.