While the Prophet model did not “nail” any price targets, I leave open to the possibility that it “may have” implied a price direction. In the previous article, code was given to generate the price target and confidence limits for the price of Facebook ($FB) shares using Facebook’s AI predictive algorithm called “Prophet”. Recapping that data from the previous article, remembering that on Friday’s close (2021-07-23) on which the prediction set was based, Facebook closed at $369.79. The prediction for the close of the following Monday (2021-07-26) was between $326.69 and $354.75 which would have been a significant loss.
|Data||Date||Regression||Lower CL||Upper CL||Upper CL||Price (Close)||% change|
|Line||Close||Close||Close||%change||vs. previous day|
|Net Change||-2.99%||Net Change||-3.65%|
Prophet is a procedure for forecasting time series data based on an additive model where non-linear trends are fit with yearly, weekly, and daily seasonality, plus holiday effects.GitHub
While $FB didn’t fall within the predicted upper and lower limits over the course of the week, it did decrease in price over the course of the week. Interestingly, whether by coincidence or not, the decrease was approximately the same amount as the predicted net change of Prophet’s upper confidence limit. Given the highly unpredictable nature of monetary markets in the long term, and in acknowledging that securities can only do one of four things (go up, go down, stay the same, or cease to exist); it would appear at first glance the probability in the decrease in price is on par with the toss of a coin but is this the case?
The only immediate way to answer this question is to continue to monitor the generalized trend and price action over time to see if there is any statistical merit to using the algorithm in this capacity, and possibly by comparing it to other models better suited for this data set (if there is such a thing). For the sake of argument here are the predicted values for the rest of this week:
It will be interesting to further model $FB when the current price is within the confidence limits of the model. As a researcher, I do caution anyone doing such analysis to do so with a contrarian view. Don’t let your bias or desire to see a pattern be your underlying decision to accept the results. For example, while $FB’s quarterly results from this past week were positive, they gave guidance that was tempered in outlook. Analysts conclude this is the reason for the drop in price towards the end of the week in question, and they anticipate a further dip. Is Prophet capable of anticipating this? There are simply too many unanswered questions at this stage to rely on any algorithm as a sole decision-making tool. Continual testing, optimization, and modeling is the only way to know. The goal in these articles is to make an attempt to evaluate the Prophet AI model over time, documenting the results here.
As always this is not trading advice, nor is it a recommendation of any sort. I am not a financial advisor nor would I pretend to be one, even if I stayed at a Holiday Inn Express. “If” there is any seasonality to the market place one would expect the following generalizations to hold true. At some point evaluating these charts with historical data might prove interesting to the self-fulfilling prophesy behind the idea that “history repeats itself”. I’ll try to do some comparisons in a follow-up.