Students are sometimes confused by the distinction between prospective and retrospect cohort studies. The key distinction is really based on when the investigator conceives the study and begins to enroll subjects relative to when the health events of interest have occurred.
But let's take a closer look at this.
Both retrospective and prospective cohort studies follow the fundamental cohort study design. They identify a cohort or group of individuals who initially do not have the outcome of interest. The cohort is then grouped based on their exposure status and both exposed and non-exposed individuals are then followed forward in time to see what subsequently happens to them in order to compare incidence and look for associations.
First, let's take a look at the prospective design. We're going to look at the Nurses' Health Study as an example.  
This was a prospective cohort study that began in 1980.
The investigators wanted to understand more about the determinants of heart disease and cancer in women, and they decided to follow female nurses, because they offered many advantages in terms of long-term follow-up.
So they began to enroll nurses, and they had the advantage here that they could verify that, in fact, they didn't have any of the health outcomes of interest at the beginning of the study,
They could also very carefully assess their baseline exposure status with respect to many possible exposures.
When they had finished their enrollment, they had identified a cohort of 117,000 nurses who all were known to be without cancer or cardiovascular disease, and they could then began to follow this cohort forward in time, checking with them periodically every year or two years when they would administer additional follow-up questionnaires that would record any changes in their exposure status, and also identify any health events, health outcomes that were of interest to the investigators.
After some time had elapsed and health outcomes had begun to accrue, they could then begin to test hypotheses. So, for example, they could take their cohort and they could group them by body mass index and perhaps compare obese and lean women to see if obesity was associated with an increased risk of heart attack.
In addition to looking at this particular exposure the investigators have the possibility of testing many hypotheses to answer many questions about a whole variety of other exposures, looking at, for example, at nutritional habits, exercise, and so forth.
The key thing here is that the study was pre-planned and pre-designed to answer specific questions.
They had the ability to verify that the subjects they enrolled, in fact, were free of the outcomes of interest at the beginning, and they could assess their baseline exposures status with respect to many any of them had developed any of the outcomes. 
So this makes it possible to greatly reduce misclassification bias, and by collecting information on many exposures at the beginning it also gives them the opportunity to control for confounding factors.
Now let's contrast this with a retrospective cohort study. Here our investigator wants to determine whether chemicals used in tire manufacturing increase the risk of death.
He's of dealing with a rare exposure, and he decides to jump back in time to identify his cohort.
He knows of a tire manufacturing company that has been in operation for several decades, so he jumps back in time to study this cohort.
Obviously, at the point in time at which he goes back to identify the subjects, they do not yet have the outcome of interest, but in the interim some of these individuals have in fact died. So, some of the health events of interest have already occurred by the time the investigator conceives and initiates this study.
Nevertheless, by jumping back several decades, he's able to identify the employees, and in this case he has to use their employee health records as his primary source of data.
This would be a good source of data for determining whether they had the exposure of interest or not. 
So, for example, individuals who were directly involved in the manufacture of tires who would have been exposed to the chemicals were clearly identified, and other employees of the company who were perhaps administrative or sales force employees would serve as a non-exposed comparison group.
Having identified these two exposure groups, one could then see what subsequently happened to them, and here, of course, the outcome of interest is death, and we can use the employee records, next of kin, and perhaps the National Death Index and other sources of information to ascertain whether or not these individuals subsequently died. 
But notice that this outcome of interest has occurred prior to the investigator initiating the study.
The other problem with this study is that the study was not pre-planned. The data that they were using was based on the employee health records, which are clearly not designed to conduct a scientific study.
They would've had many omissions in data, and this would've made it difficult for the study. For example, there likely would have been some misclassification of various a status exposures, and there may have been a lot of missing information with regard to many other exposures that can be confounding variables.
So, this study has a lot of disadvantages compared to the prospective design.
This cartoon, summarizes the key difference between these two designs. In a retrospective design the investigator jumps back in time to identify the disease-free cohort and then looks at what subsequently happened to them, but the health events of interest occurred prior to the initiation of the study. In contrast, with the prospective design the study is pre-planned. 
The investigator can enroll individuals who are disease-free or do not have the health outcomes of interest initially, and they can very carefully design data collection instruments that will allow them to assess many baseline exposures prior to anybody developing any of the outcomes of interest. This offers very distinct advantages in terms of minimizing bias and also providing better control of confounding.