By assuming the procedure you followed in replicate the model is correct and there are not errors in data mining or quality, your findings could be affected and influenced by several reasons.
I report as follows those that, according to me, could be the main ones:
- Data sample: the dataset you used to replicate the Fama-French model could be too little in terms of observations and time-period; keep in mind that in almost all their papers, Fama & French used a data sample that allows to analyze several years (about at least 20 years).
So, try to increase the time period and the number of observations of the dataset to get better results;
- Portfolio coarseness: some of your portfolios could be composed by too little stocks and the results could be biased; the factor loadings should be monotonically decreasing in the market capitalization.
Be sure that each portfolio is composed by the same number of stocks (more or less), constructing them by using the distribution percentiles.
Populated class: each portfolio has to be populated enough with a minimum number of stocks that have more or less similar economic features and eliminate the others.
Econometrics issues: check that all hypothesis about the linear regression model are fulfilled (Normality, Homoskedasticity, autocorrelation,...) in order that your results ac
quire more reliability.
Time period: may be that their results are influenced by the time period they used in their analysis and that the phenomenon disappeared during the last years; the SMB variable is based on the Small Size effect and, it is proven that this kind of phenomena disappear after someone makes them of public domain, e.g. publishing papers. Think about the January effect: there are a lot of papers that documented it disappeared!
Those are some of the main issues could influence your results.
Check them and after, if they are all satisfied, read the later Fama & French's papers to check if their way to analyze the sample is equal to the yours.