Weather prediction is difficult.
Most data analytics courses only predict past data and use train/test sets. Predicting real events which have yet to occur is far more difficult and exciting. The learning that comes from genuine prediction will put you far ahead of peers.
We are going to attempt to predict the weather. Weather prediction is a great way to learn data analytics because it contains continuous predictions (temperature), discrete predictions (rain/no rain), and has many possibilities for using unsupervised techniques (like clustering and visualization). Weather predictions require careful thought about how to structure data. The data is public and comes in daily, which makes evaluation easier.
The National Oceanic and Atmospheric Administration (NOAA) publishes hourly and daily weather observations from weather stations around the world.
The GSODR R package makes it easy to
download the latest weather data. We will use the Albany Internation Airport
as the location for the weather data and forecasts. The station ID is 725180-14735.
Since the course is using Python, and I have not found a suitable Python
replacement for GSODR, I will post the historical forecast data to
Slack using a bot. You can either use the data from Slack, or run the
R script download_latest_data.R.
You should submit the forecasts for the daily high temparature (MAX),
daily low temperature (MIN), the probability of precipitation (P_PRCP). Preciptation
will be have occurred when the value of the PRCP column is greater than 0.00.
The baseline prediction is named mean_prev_seven. It is simply the average temperature
from the last seven days.
The submissions must be in a CSV file formatted exactly as the
values in the tables below. The easiest way to acheive this
format is to build your forecasts with the baseline_predictions.py
file. Modify the code in that file to include your forecasting technique.
| fc_date | fc_name | fc_var | fc_value |
|---|---|---|---|
| YYYY-MM-DD | your_name | VAR_NAME | Degrees (C) |
Example:
| fc_date | fc_name | fc_var | fc_value |
|---|---|---|---|
| 2019-02-04 | lspitzley | MAX | 5.5776 |
| 2019-02-04 | lspitzley | MIN | -0.5468 |
| 2019-02-04 | lspitzley | P_PRCP | 0.2556 |
A one-week forecast will should have 21 rows.
You can test your format by running your forecasts through the
forecast_evaluation.py script.
Continuous forecasts are evaluated with
sklearn regression metrics.
Binary predictions are scored with standard methods that we covered
in class.
The forecast_evaluation.py script reports scores to the console and to
a .csv file in the output/ directory.
The RMSE will be the primary scoring metric for the temperature predictions. Log loss and AUC are the primary metrics for precipitation predictions.