This post is about a predictive analytics tool (in beta version) I started building during the first lockdown in March 2020. I used R and Python for this purpose, and more specifically Flask and rpy2. It’s been a pretty cool and instructive experience for me, of having both R and Python interacting into a web app, and deploying the tool on Heroku:

https://www.techtonique.net

What’s inside? It’s a suite of Application Programming interfaces (APIs), a system which can receive requests from your computer keyboard, to carry out given tasks on given resources. This system is programming-language-agnostic (works for curl, Python, R, Javascript, PHP, etc.), relatively fast, and there’s no need for additional packages installation when you have an access to it.

What are requests and resources here? In Techtonique/APIs, resources are Statistical/Machine Learning (ML) model predictions.
In this context, one type of request for a task could be: obtain sales, weather or revenues forecast for the next 5 weeks.
In general though, requests for tasks do not have such a long description. They are characterized
by a verb and an URL Path – approximately, an endpoint – which both lead to a response.

Below in an illustrative example of API. In this example, the resource we’re interested in is a list of users that we’d like to manage.

Request type (verb): GET
    URL Path: http://users  |  endpoint: users  |  API Response: display a list of all users
    URL Path: http://users/:id  |  endpoint: users/:id  |  API Response: display a specific user

Request type (verb): POST
    URL Path: http://users  |  endpoint: users  |  API Response: create a new user

Request type (verb): PUT
    URL Path: http://users/:id  |  endpoint: users/:id  |  API Response: update a specific user

Request type (verb): DELETE
    URL Path: http://users/:id  |  endpoint: users/:id  |  API Response: delete a specific user

In Techtonique/APIs, a resource endpoint will be for example /MLmodel instead of /users. And since the resources are already implemented – as model predictions – and do not need to be altered (PUT) of deleted (DELETE), every request for a forecast is a POST request to a /MLmodel.

So, how does the tool work? You start by creating an account (sign up @ https://www.techtonique.net) using a valid email address. Then, prepare an input file containing univariate or multivariate time series, with the following format.

The first column of the input csv file must be named ‘date’. In addition, the dates in this column must be regularly spaced, and have the format ‘%Y-%m-%d’ (year, month, day, separated by -‘s). The rest of the procedure is described below for curl, Python and R (in that order). For more details on models’ parameters, log in @ https://www.techtonique.net.

With curl

Obtain a token with the credentials that you used for signing up @ https://www.techtonique.net:

curl -u yourusername:yourpassword -i -X GET  https://www.techtonique.net/token

Use the token for obtaining dynrm forecasts. dynrm is adapted from NNETAR, but uses
an automatic ridge regression to adjust time series’ lags and seasonal lags.

curl -u token:x -F 'file=@/path/to/univariateinput.csv' " https://www.techtonique.net/dynrm?h=10&type_pi=A"

With Python

ridge2 (cf. second example) is the model from Moudiki et al. (2018) for multivariate time series.

import requests


base_url = "https://www.techtonique.net" 

# same credentials that were used for signing up @ https://www.techtonique.net
username = "yourusername"
password = "yourpassword" 


# 1 - request for a token -----

response_token = requests.get(base_url + '/token', auth=(username, password))

token = response_token.json()['token']

print("\n")
print(f"token: {token}")


# 2 - request using a token (univariate time series) -----

# for more details on these parameters, log in
params = (
    ('date_formatting', 'ms'),
    ('h', '5'),
)

files = {
    'file': open('/path/to/yourinputfileunivariate.csv', 'rb')
}

response = requests.post(base_url + '/dynrm', 
params=params, files=files, auth=(token, 'x'))

print(response.json())


# 3 - request using a token (multivariate time series) -----

# for more details on these parameters, log in
params = (
    ('date_formatting', 'original'),
    ('h', '8'),
    ('nb_hidden', '10')
)

files = {
    'file': open('/path/to/yourinputfilemultivariate.csv', 'rb')
}

response = requests.post(base_url + '/ridge2', 
params=params, files=files, auth=(token, 'x'))

print(response.json())

With R

library(httr)
library(datasets)


# same credentials that were used for signing up @ https://www.techtonique.net
username <- "yourusername"
password <- "yourpassword" 


# 1 - request for a token -----

res_token <- httr::GET(url = 'https://www.techtonique.net/token',
                 httr::authenticate(username, password))


(token <- httr::content(res_token)$token)


# 2 - request a forecast using a token -----

# for more details on these parameters, log in
params = list(
   `date_formatting` = 'original',
   `h` = '15'
 )

files = list(
   `file` = upload_file('/path/to/AirPassengers.csv')
)

(res_api <- httr::POST(url = 'https://www.techtonique.net/dynrm', query = params,
                  body = files, httr::authenticate(token, 'x')))

list_res <- httr::content(res_api)


# 3 - Extract results -----

h <- length(list_res$ranges)

forecast_object <- list()
forecast_object$mean <- forecast_object$upper <- forecast_object$lower <- rep(0, h)

for (i in 1:h)
{
  forecast_object$mean[i] <- list_res$averages[[i]][[2]]
  forecast_object$lower[i] <- list_res$ranges[[i]][[2]]
  forecast_object$upper[i] <- list_res$ranges[[i]][[3]]
}


# 4 - Plot results -----

par(mfrow=c(1, 2))
plot(AirPassengers, main="Dynrm forecast")
xx <- c(1:h, h:1)
yy <- c(forecast_object$lower, rev(forecast_object$upper))
plot(1:h, forecast_object$mean, type='l', main="Dynrm forecast \n using Techtonique's dynrm")
polygon(xx, yy, col = "gray90", border = "gray90")
lines(1:h, forecast_object$mean, col="blue")

You may experience some requests timeouts from time to time or a small volatility in API calls’ durations (choice of a low cost infrastructure for now, working on adjusting it in this beta version). When/If it happens: just try again. Feedbacks/suggestions are welcome as usual. For this specific web application, please use: [email protected]