Prophet stock prediction

Prophet library has been developed by Facebook as a tool for time series forecasting.

Neural network machine learning models for time series prediction - like LSTM neural nets need typically thousands of instances of time series to learn from. On the other hand, Prophet learns only from the one time series that is given to it and makes predictions only based on that. That is great benefit. For the usecase that has been tested here the learning and predicting part takes only few seconds per one time series instance.

Super cool and super simple.

Prophet works incredibly well on historical data that has reocurring pattern - some kind of seasonality. Yearly, monthly, weekly or daily seasonality.

Hence Prophet library is ideal for prediction of electricity consumption during the year or sales of already established product with long selling history.

But what about stocks?

Stock price is not seasonal in general. While some sectors are cyclical, in general we cannot say that every Tuesday Google stock goes up.

Making predictions with prophet is super easy, prediction takes only one line of code, all the heavy lifting is done under the hood.

Prophet().fit(df)

Of course, that Prophet library does not know the future. But it learns from past dataset and past priceaction and tries to extrapolate the behaviour into the future.

Example prophet predictions - VNO and XOM stocks

prophet_stock_prediction.png

prophet_stock_prediction_XOM.png

Above are examples of two stocks with rather complex behaviour (they just dont go up or just down). Looks like Prophet learned from the steep drop and in further predictions also predicts another drop after significant price runup. That intuitively seems like reasonable thing to do.

Import libraries

In [51]:
# Optional installation with Anaconda
# conda install -c conda-forge fbprophet
In [52]:
import pandas as pd
from fbprophet import Prophet

from datetime import datetime
import yfinance as yf
In [53]:
#suppress warnings
import warnings
warnings.filterwarnings("ignore")

Actual predictions

We will download daily data from yahoo finance and make prediction n days into the future. Prophet requires us to use specific names for for time(ds) and price columns (y).

# prophet needs specific column naming
df['ds'] = df['Date']
df['y'] = df['Adj Close']

So we just feed the data into prophet and make predictions.

See more info on official Prophet quick start guide: https://facebook.github.io/prophet/docs/quick_start.html

In [54]:
start=datetime(2018,1,1)
end=datetime.now().date().isoformat()    # today

# pick stock ticker
# symbol = 'BTC-USD'  # usage for crypto
symbol = 'AAPL' # usage for stocks
In [55]:
df =  yf.download(symbol, start=start, end=end)

[*********************100%***********************]  1 of 1 completed
In [56]:
df.head(2)
Out[56]:
Adj Close Close High Low Open Volume
Date
2018-01-02 41.380238 43.064999 43.075001 42.314999 42.540001 102223600
2018-01-03 41.373032 43.057499 43.637501 42.990002 43.132500 118071600
In [57]:
df = df.reset_index()
In [58]:
df.head(2)
Out[58]:
Date Adj Close Close High Low Open Volume
0 2018-01-02 41.380238 43.064999 43.075001 42.314999 42.540001 102223600
1 2018-01-03 41.373032 43.057499 43.637501 42.990002 43.132500 118071600
In [59]:
# prophet needs specific column naming
df['ds'] = df['Date']
df['y'] = df['Adj Close']
In [60]:
df.head(2)
Out[60]:
Date Adj Close Close High Low Open Volume ds y
0 2018-01-02 41.380238 43.064999 43.075001 42.314999 42.540001 102223600 2018-01-02 41.380238
1 2018-01-03 41.373032 43.057499 43.637501 42.990002 43.132500 118071600 2018-01-03 41.373032
In [61]:
m = Prophet()
#m.fit(df.iloc[:-365])
m.fit(df)

INFO:fbprophet.forecaster:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
Out[61]:
<fbprophet.forecaster.Prophet at 0x7fb254e63710>
In [62]:
future = m.make_future_dataframe(periods=180)
future.tail(2)
Out[62]:
ds
974 2021-08-28
975 2021-08-29
In [63]:
forecast= m.predict(future)
forecast.tail(2)
Out[63]:
ds trend trend_lower trend_upper yhat_lower yhat_upper additive_terms additive_terms_lower additive_terms_upper multiplicative_terms multiplicative_terms_lower multiplicative_terms_upper weekly weekly_lower weekly_upper yearly yearly_lower yearly_upper yhat
974 2021-08-28 172.986227 155.035460 192.657953 165.407147 204.531512 11.214497 11.214497 11.214497 0.0 0.0 0.0 -0.510364 -0.510364 -0.510364 11.724861 11.724861 11.724861 184.200723
975 2021-08-29 173.176393 155.039521 193.107103 166.144027 203.887040 11.196193 11.196193 11.196193 0.0 0.0 0.0 -0.510364 -0.510364 -0.510364 11.706558 11.706558 11.706558 184.372586

Built-in Prophet plot

In [64]:
fig1=m.plot(forecast)

Custom plot with matplotlib

Since we have the dataframe generated by Prophet, it is very easy to plot it with our favourite python plotting library. In this case we are utilizing matplotlib.

In [65]:
import matplotlib.pyplot as plt
In [66]:
plt.figure(figsize=(15,5))
plt.title(symbol + ' stock price prediction')
plt.plot(df['ds'], df['Adj Close'])
plt.plot(forecast['ds'], forecast['yhat'])
plt.plot(forecast['ds'], forecast['yhat_lower'], alpha=0.1, color='blue')
plt.plot(forecast['ds'], forecast['yhat_upper'], alpha=0.1, color='blue')

plt.fill_between(forecast['ds'], forecast['yhat_upper'], forecast['yhat_lower'], alpha=0.1)
#plt.plot(forecast['ds'], forecast['trend'])

plt.show()

Sources:

Facebook Prophet documentation