Automated Time Series Analysis: Forecasting with No Code Machine Learning

Automated time series is now available from Squark.  What is that, you ask?  It is a forecast of future values over time generated from past data.  Think of how the price changes every day for your favorite stock. Time-series forecasting is being able to predict the price of that stock over multiple time periods.  For example, you may want to forecast what Tesla’s stock price will be for the next 60 days or across other time periods (for example, the next 360 seconds or the next 8 quarters).  Other examples of time-series data are the weekly numbers of account signups, your daily revenue, your hourly transactions, and so on.  Any number with a date associated over time is a time series.  And now, any time series can be predicted with Squark: seconds, minutes, hours, days, weeks, quarters, months, and years. 

Automated time series by Squark isn’t simply a new type of regression.  Squark has always, since its launch, automated many types of machine-learning regression.  While people can consider regression a type of time series analysis, let’s face it, regression is not true time series.  Regression predicts one number.  It is awesome when we predict a single number, like “Next Week’s Sales” or “CLTV,” but how is time incorporated into regression?  Usually, it is not!  In Squark, we embed a factorization capability which allows us to deconstruct dates into their core elements (like day of the week, etc.).  Squark regression thus uses date factors, but factoring is not a true time series analysis (as your college professor would say). 

Squark now has true time-series automation in our SaaS using our massively parallel GPU-based architecture (thanks to NVIDIA). And don’t forget we can install Squark on bare metal, Docker, Kubernetes, Lambda, or even export the model code. No one else has these advanced capabilities but Squark. And no one else will provide it so elegantly, visually, simply, and economically at such a performant scale.

Automated Time Series with Machine Learning: Business Uses Cases and Value

1. Analytics, data science, insights, and finance teams across every industry can benefit from automated time series forecasting in Squark.  The analyst simply uses Squark connectors to access data wherever it exists (or use our API).

2. Forecast the future movements of your key performance indicators (KPIs), such as net retention in SaaS, subscription, and D2C businesses.

3. Identify future periods of bearish or bullish performance.

4. Perform what-if analysis across data meaningful to your business outcomes.

5. Predict the impact of business actions today on future performance tomorrow.

6. Extend your dashboards (in whatever BI tool you use – Domo, Tableau, PowerBI) to include accurate forecasts.

CMO’s can forecast ROI, ROMI, ROAS.  CFO’s can forecast revenue, costs, and truly any number within the world of finance. Analytics leaders can firm up their grasp of predicted future KPI performance. Product managers can forecast the uptake of their product and the usage of key features of time. Operations and IT teams can de-risk performance by understanding what may happen over time. Retailers and eCommerce companies can forecast transactions and inventory.  Sales teams can forecast wins and revenue.  And of course, all this forecasting can be done in parallel and with deep detail across geographies, business units, and the meaningful segments in your business.

Automated Time Series with Squark:  Type of Time Series Analysis Possible

Squark uses our proprietary automated AI and machine learning platform to deliver the following time series models:

1. Univariate time series. Forecast one value from a set of dates.  This amplifies the pattern in existing data.  For example, a univariate time series analysis would forecast how many accounts will sign up over the next 30 days by day.

2.  Multivariate time series. Forecasts one or more values from a series of values and a date.  If you wanted to forecast the per-minute stock price for 10 stocks, for the next 360 minutes, you would use multivariate.  Or, if you wanted to forecast your conversion rate, revenue, users, and sessions by day, you would use multivariate.  Multiple numbers can be forecast with a mathematical understanding of how the other numbers impact each other.

3. Multimodal time series.  Forecast one or more values from a series of values, a date, and categorical variables. For example, say you are Walmart, we can include data about the weather or your product categories along with multiple time-series numbers and use all of that information as part of the forecast.

Automated Time Series with Squark:  How It Works

Our team quickly realized that the approaches traditionally used for automating time series were wholly insufficient for the job.  Many libraries may create a time series model, but that model cannot be applied to new data for actually forecasting the future time series.  Thus, we built our own learning algorithms for time series using TensorFlow as well as expanded and made better FBProphet, SARIMA, ARIMA, VAR, and other algorithms too (ask us!).  For all these algorithms, it is important (for data scientists) to know that we fully explore and tune all the time series hyperparameters, not with some random search, but with our proprietary approaches for full exploration across hundreds of thousands of hyperparameters. The result is the best auto-tuned time series model (out of thousands we create) that money can buy.  All done in minutes, thanks to our massively parallelized GPU architecture.

Squark’s Automated Time Series has Powerful Pre-Processing and Feature Engineering

Automating feature engineering is an entirely different beast in time series too.  Our automated feature engineering in classification and regression wasn’t able to be fully applied to this new model class, so we built out an amazingly rich capability for auto-feature engineering time series data.  For example, it is important to be able to determine distributions and statistics, moments, outliers, gaps, duplicates, extremely granular time periods, and many other data conditions and oddities for time series machine learning.  So, we built that proprietarily into Squark.

Performance is stellar too. Predictive results render in minutes, and, like all Squark, you then operationalize via sending the data back to another system, exporting the model code, calling our API, or maybe you want to deploy on-prem?  Whatever your modality to operationalize time series, Squark’s automated time series has you covered for MLOps.

Squark is today’s powerful automation of AI that we have abstracted into clicks for analysts and data scientists while creating a system bought by the business because it has a high ROI. Squark scales to the biggest of data, automatically connects, discovers, preps, engineers, models, explains, simulates, and activates machine learning predictions from the Fortune 100 to the mid-market.  Squark’s features and capabilities go far beyond what you will find available in code or in other tools. Try out Squarks’ automated time series for yourself by reaching out for a demo.