Improve ML-Powered Time Series Forecasting with PredictHQ

How to improve time series forecasting

New methods to improve analysis and forecasting accuracy

Time series forecasting is a quantitative approach that uses information regarding historical values and associated patterns to predict future observations such as demand. Often, it’s broken down into trend analysis, seasonality, and irregular components. 

• Trends: Trend data provides insight to the long term direction of predicted demand based on customer interest and adoption.

• Seasonality: Patterns of systematic demand shifts occurring at specific, regular intervals are recorded and denoted as seasonal variations.

• Irregular components: The unsystematic fluctuations caused by external factors creates volatile demand. 

Most demand forecasting practitioners have access to historical time series data that can be used to understand baseline demand like trends, seasonality, and cyclical patterns. By combining statistical approaches with machine learning, time series data can be analyzed and used to generate demand forecasts. 

But how do your time series models factor in irregular components like volatile demand? The unsystematic fluctuations create demand anomalies which generate incremental demand: a surplus of demand for a business or decremental demand, diminishing demand for the business. It’s challenging for teams to track demand anomalies, pinpoint why they occur, and train their time series models. But doing so is crucial to improve time series forecasting accuracy. 

Predictive demand intelligence provides breakthrough context and visibility into demand anomalies. Events like sports games, health warnings, and school holidays are demand causal factors that are major drivers to both incremental and decremental demand. With our demand intelligence API, you can access millions of verified external factors that cause anomalies.

We include features in our data that can help explain demand anomalies and prioritize what to incorporate into your models. Some features include:

• Ranking: We’ve built a ranking system to identify the impact of events. Our ranking algorithms are optimized to highlight event impact on different industries including retail, food service, aviation and more.  

• Event categories: PredictHQ covers 19+ major event categories world wide making it easier to identify what’s driving demand anomalies.

• Labels: Once our machine learning models sort events into categories, NLP models add labels to every event so you can build hyper-targeted models.

Use predictive demand intelligence to identify what’s driving demand anomalies, and train your models with this new data set to improve time series forecasting. Advanced knowledge of external factors means you can tailor your pricing, supply chain, labor, and marketing strategies to unlock profits and create operational efficiencies.

Enhance time series methods and ML techniques with demand intelligence

No matter the method - you need accurate, reliable data

Time series forecasting is full of quantitative methods and techniques. Some of the commonly used statistical techniques of forecasting include moving average based methods like Autoregressive Moving Average (ARMA) and Autoregressive Integrated Moving Average (ARIMA). In recent years, machine learning methods like Long Short-Term Memory Networks (LSTM) have been proposed for the task of time series forecasting as well. The common objective of all of these techniques is to improve forecast accuracy by minimizing the difference between predicted and actual demand values. 

But improving forecasting accuracy with these statistical methods and machine learning techniques can’t be achieved without accurate, reliable data sources that are optimized to forecast external demand causal factors. 

Think about an annual business conference that changes city location each year or sporting teams that suddenly gain traction in a particular season. These changes are hard to track but impact the strength of demand causal factors which will impact demand forecasts. PredictHQ is able to generate, standardize, and store this data by using our extensive knowledge graph which includes metadata about venues, sports, teams, performers, and more. With this, our systems are able to track events that could be causing your demand anomalies. Additionally, demand intelligence is designed to quantify event impact both at the individual and aggregated levels so demand forecasting practitioners can utilize these features into their time series forecasting models. 

PredictHQ provides the APIs and tools to use in forecasting so you can model the impact of demand causal factors. By following the demand forecasting guide, you can learn how to add features into your model and fine-tune demand prediction precision. 

Ensure data quality and improve forecasting accuracy with our pipeline

Improving prediction precision starts with the data

Turning data into valuable, reliable intelligence is a labor-intensive challenge that many data teams try to tackle each day in order to improve. In fact, data scientists spend 40% of their time gathering and cleaning and only 20% of their time on building models. Our data processing systems aggregate and verify millions of causal factors to ensure data quality. Now data science teams can stop wasting time on the intricacies of data cleaning/standardization and focus on analyzing time series data and putting forecasting models into production to create business value.

We enrich data with entity information, geocode data and other predicted attributes so teams can leverage demand intelligence.  Ranking events by predicted impact is a key part of our data enrichment process. By using our ranking technology, forecasting models can prioritize demand data based on anticipated impact.   

This complex pipeline delivers data intelligence that you can easily integrate into time series forecasting. Correlating your historical data to demand intelligence is key to discovering which events drive your business. Once you establish correlation, use PredictHQ future data to train your machine learning models and update your forecasts. 

Data teams have to validate and test their time series models and can look at time series cross validation. This procedure requires a series of test sets, each consisting of a single observation. Forecasting accuracy is computed by averaging over the test sets.