Read part 1 here.
Challenges in demanding forecasting
One of the biggest challenges faced by business executives for accurate demand forecasting is demand volatility. In addition to volatility as digital transformation goes mainstream in enterprises large and small, companies are grappling with multiple channels of demand.
In manufacturing sector, product demand could originate from multiple channels like re-
sellers, e-commerce sites, direct consumers, sales etc. This is causing fragmented demand signals resulting in inaccurate forecasting.
Several small to medium sized businesses rely on rudimentary forecasting tools that rely on incomplete data or forecasting models that are static and do not consider for example promotions by competitors or new products with no historical data for forecasting.
There are too many factors influencing demand, ranging from weather fluctuations to posts by social media influencers, causing customers to frequently change their minds causing demand fluctuations.
Lack of macro economic view
Majority of enterprises lack mechanisms to adjust forecasting based on consumer/customer purchase behavior, macro economic conditions and factors influencing demand in real-time.
Developing more advanced forecasting techniques involves knowledge of technologies like AI/ ML, advanced statistics and real- time data analysis.
AI augmented demand sensing
Most companies base their forecasting on historical data from internal sources like sales channels, production systems, inventory etc. However as Covid-19 has shown, relying purely on internal sources and not considering external factors produces forecasts that are not realistic and error prone. Demand sensing incorporates external data like weather, consumer sentiments, competitors product promotions etc. and applying AI on top of these demand signals.
Benefits of AI driven demand sensing
- Digitization of supply chains
- Ability to analyze structured and un-structured data
- Increases supply chain visibility and agility
- Automation capabilities
- Improved accuracy due to macro and micro analysis
- Anticipates customer behavior in real-time
- Captures short-term demand signals
- Breaks various silos and fosters easy data availability as needed
8 steps for improving forecasting accuracy with AI
All the sophisticated tools to influence demand with pricing, new product introductions and promotions have pushed demand volatility and consumer expectations to unprecedented levels. To stay competitive and keep costs low, companies should augment existing statistical techniques with AL/ML and deep learning models.
1. Extract data from multiple sources
Ingest data from multiple relevant internal sources like PoS, Sales forecasting, production planning. If there is a distributor/retailer in the loop ingest their inventory levels as well and any additional external data for macro level analysis.
2. Data Pre-processing & Feature Engineering
With data potentially coming from multiple sources, it needs to be cleansed and correlated using data analytic and data exploratory tools.In addition using techniques like principal component analysis (PCA) number of features should be reduced for better forecasting model accuracies.
3. Classify products and use right forecasting method
If forecasting new products there will not be prior data available for forecasting. In such scenarios use qualitative methods outlined earlier. If forecasting existing products or products similar to existing ones use quantitative methods.
4. Define forecast period
Define the forecast period depending on the objectives. If the goal is for inventory replenishment a shorter time horizon like 60-90 days is ideal whereas for production planning longer time periods like 12 months is better.
5. Create a rudimentary baseline forecast
Create a rudimentary forecast that acts as a baseline for comparing forecast models against more advanced Time Series and Deep learning models. Simple methods like moving average calculation or exponential smoothing (if seasonal) will suffice.
6. Create advanced forecasting models and compare against baseline model
Using AutoML tools and historical data (training data), create multiple forecasting models with Regression, Classification, Deep Learning and time series algorithms. Measure the performance of each forecasted model and compare against the baseline model. Discard models that do not perform well compared to baseline models.
7. Create an integrated (ensemble) model from multiple models
Consider the performance of each model in time and combine the weighted predictions of the best performing models from various approaches like simple statistical models to Deep Neural network based models for a more accurate demand forecasting model.
8. Track real-time demand data against forecast models and train models regularly
Track the actual demand with the forecast time period granularity (ex: weekly/monthly etc.). Train the forecasting models with data to improve the accuracies of the predicted models. Continue adjusting the production planning based on updated models.
How Predactica helps with demand forecasting for enterprises
Accurate forecasting is key to optimizing inventory and improving profitability. This requires enterprises to be able to adjust forecasting in real-time in a cost effective manner to reflect demand in the field. Predactica’s platform meets these requirements through a cloud native platform built from ground up to enable companies of all sizes and verticals to leverage the power of AI/ML technologies.
Consumption based SaaS pricing
With Software as a service (SaaS) model and consumption based pricing, users only pay for actively used compute resources. No more paying for under utilized subscription licenses reducing operational costs.
No-code data science platform
With intuitive end-to-end data science capabilities, citizen data scientists can create powerful and accurate predictive models in minutes without writing a single line of code.
Advanced forecasting techniques
In addition to statistics based forecasting models, Predactica supports AI and deep learning forecasting models resulting in improved forecasting accuracies.
All functional capabilities of Predactica are exposed as ReST APIs. This helps customers integrate the forecasting data predictions with downstream production and inventory planning applications.
Real-time data processing
Using efficient in-line processing functionality, data transformations are done in real-time enabling creation of accurate forecasting models.
In industries like Retail, CPG & Manufacturing, demand forecasting is one of the main problems facing supply chain teams to optimize inventory, reduce costs, and increase sales, profit, and customer loyalty. To overcome this issue, there are several methods such as rudimentary statistical methods, time series analysis and machine learning approaches to analyze and learn complex interactions and patterns from historical sales data from multiple channels.
By enriching the input data for forecasting from a combination of external sources like economic research, consumer trends along with internal data sources like Point of Sale (POS), sales pipeline and inventory data enterprises will have a 360 degree view of the demand increasing the accuracy of forecasted models.