Well, assuming that you did everything correctly in technical matters and there are no software bugs, etc., the only viable explanation is that your data does not support accuracy higher than 70%. Did you curate your data prior to modeling? "Curate" here means careful and thorough check of the reported data values. In my field - chemistry - this would mean downloading original publications, verifying the numbers, checking correctness of molecular structures, checking consistency of measurement units, etc. A lot of tedious work, in other words.

Now, you can check the amount of consistency in your data, its "modellability", by clustering the 200 objects by similarity in the feature vector space. Do most similar objects belong to the same classes? If not, then you have a problem: classifiers working in that space will not be able to resolve them. Do the outlier analysis in the same context.

Importantly it depends on how much training data you have for every class. If you have more classes then it is possible that you have less data to specify each class for your  classifier to differentiate among them.

Another possibility is maybe your features doesn't fit enough to get better classification. Many factors can effect this like if data is not normalized for instance. 

An empirical test to check if your features separates your classes well: extract the first two or three principal components and display them on a 2D or 3D scatter plot (use a different marker for each class). That will also help you see whether there are outliers that may affect the results.

There can be different reasons as, for example, some input features are "noisy" and "non-infornative" (useless in separation of classes that You have); training sample is not large enough; a classifier used is not good enough for Your goal. Although I agree with recommendations given above by other people, I'd like to say that it is difficult to find a reason (or several reasons) does not work well without explaining a task that You are trying to solve more in detail. 

By 200 data, I assume you mean 200 training instances.  This is a small amount given that you have a vector of more than 150 features.  One likely problem is the "curse of dimensionality" - high dimension means a very large space that's difficult to search.  If possible, try getting more training instances and rank the feature set using PCA (as described above), ANOVA, or some other method.  Once ranked, try using a subset of only the strongest features - start with 20, then 30, then 40, etc.  Reducing dimensionality may help. Also, as suggested above, 70% might be the best you can do with your data.  In some cases, especially multi-class problems, 70% can be pretty good.  Good luck.

classification results are often affected due to overfitting. It means that there are no variations among the training samples among a class of objects. Try to have variation with every training sample that is used.

Hi Sara. You are not telling us how many classes you have. Do you get similar performance on all of your classes? 

I agree with Gene, for a vector size of 150, 200 samples are not enough. Mainly if you are doing a crossover approach.

Hi Sara. There are numerous reasons which affect the classification accuracy. In your problem, I think you haven't got enough samples in your data set in order to fill the parameters of your model. What is your maximum value for each attribute? I assume that each attribute in your data set takes 2 values(either True or False). Hence, we have 300 parameters in your model, but you have got 200 sample.

I will suggest you increase the size of the sample. How many predictive classes are in your data? You may need to check this.