ML forecasting

ML forecasting

Navigation by topic

augmento

FORECASTING

PREDICTIVE ANALYTICS

BUSINESS CONSULTING

While doing automation and business consulting in Planning, it became clear that the time had come for advanced ML-enabled forecasting solutions to replace widely used statistical methods.

This led to the creation of the Augmento business unit which employs data scientists, data engineers and business process consultants.

We have a very specific work domain. We develop forecasting models using machine learning algorithms in the area of demand and promotion forecasting, as well as help adjust your business process to accommodate a ML-enabled forecast.

Machine Learning has opened up a new era of forecasting. It gears businesses up for management by exception when time and resources are focused on what makes a real difference. 

OUR SERVICE

Solutions
  • Machined Learning-enabled demand and promotion (volume and budget) forecasting models
  • Technical support of the model after go-live (business as usual)
  • Business process adjustment in line with the new technology
Deliverables
  • Forecast accuracy growth
  • Forecast bias decrease
  • Less resource intensive forecasting
Technology
  • Multivariate Machine Learning algorithms:
– Gradient Boosting– Neural networks– Random Forest– ArimaX– Ensemble
  • Innovation
    sales forecast
  • Promo investments
    forecast
  • Baseline sales
    forecast
  • ML
    Forecast
  • Optimal pricing
    forecast
  • Off shelf
    forecast
  • Promo-sales
    forecast

A NEW ERA OF FORECASTING

Univariate

Statistical models

Forecast based on a single factor – historical sales

Highly dependent on expertise and availability of responsible people

angle
angle
angle

Multivariate

Machine Learning algorithms

Process any available internal and external data sources to find the best correlations and build an optimal forecast

Managing by exception when it comes to forecasting

FORECAST BASED ON BIG DATA

Actual sales

Trade terms

Promotions

Assortment movements

Customer stocks

Weather

Store openings

Market data

Pricing 

Media spend

Other data to feed

Competitive moves
Macroeconomic indicators
Holidays

CORE STAGES OF LAUNCHING AN ML FORECAST

  • 1

    Data
    discovery & preparation

    ~ 4-7 weeks
  • 2

    Model
    and process
    building

    ~ 8-15 weeks
  • 3

    Integration
    and go live

    ~ 3-5 weeks
  • 4

    Technical
    support

    Indefinite

HOW TO FURTHER WORK WITH THE ML FORECAST

THE ML FORECAST
Exporting an ML forecast into Excel
Integrating a ML forecast into a planning system like Anaplan, APO etc.
Visualizing a ML forecast in a BI platform like Domo, Tableau etc.
THE ML FORECAST
into Excel
system like Anaplan
BI platform like Domo

ACHIEVABLE RESULTS AFTER LAUNCH

  • ~20%
    Increased Forecast accuracy
  • 50%
    By up to 50% time Efficiency Improvement for planners
  • S&OP
    Simplified Process

    • CORE STAGES OF LAUNCHING THE MODEL

    1

    Data & Process discovery

    Investigation into all input data available to build the model as well into the current Forecasting and S&OP processes to fit in. Alignment on Project goals and KPIs before the start.
    2

    Causals engineering

    Selection of relevant data inputs (causals) to start building algorithms. Data pre-processing to feed into the model.
    3

    Algorithms development

    Writing scripts for chosen multivariate ML algorithms like Gradient boosting, Random Forest, Neural network etc. Training and testing for stability, bug fixing.
    4

    Algorithms tournament

    “Tournament” – the model picks up the most suitable algorithm for each forecast item so in the end there is an ensemble of various algorithms which make up an integral model.
    5

    Product segmentation

    ML-generated forecast is split into 3 core segments: No touch, Light touch и Heavy touch depending on the need to manually adjust the forecast.
    6

    Frontend visualization

    The model output can be displayed in Domo or any other BI platform via a special dashboard for planners and S&OP managers as well as for business users.

    Архитектура решения

    Your browser does not support SVG
    Your browser does not support SVG

    Архитектура решения

    Нажмите на иконку

    ACHIEVABLE ROI OF THE PROJECT

    If Forecast accuracy grows by + 10%

    Productivity growth

    Less manual work:

    • Less time to collect data
    • Less errors and time spent to correct them
    • Less IT involvement
    • Less time for forecasting

    Higher forecast accuracy => more reliable P&L => higher quality management decisions 

    More accurate long term forecast => more sustainable production facilities utilization => higher quality CapEx decisions

    Higher customer satisfaction: better contract terms, less penalties, stronger negotiation power

    Costs optimization

    Less out of stock:

    • Higher service level + 3%’
    • Higher revenue +1%’ (a conservative estimation for highly efficient business)

    Less overstock

    • Optimized stocks – 25%’
    • Less write-offs- 17%’

    Leaner production

    • Higher OEE (by up to 11% due to a more forecastable production schedule’)
    • More sustainable and efficient labor management

    ‘Based on an internal assessment by Advanced

    CUSTOMER SUCCESS STORIES

    FREQUENTLY ASKED QUESTIONS

    1

    Why is an ML forecast better than a statistical one?

    A statistical forecast is built using sales history only, whereas an ML forecast takes in additional valuable data influencing the demand such as promo investments, inventory, weather, etc. Modern ML algorithms like neural networks are able to detect non-linear correlations between sales and various demand drivers which no traditional statistical methods or experts can detect. Leveraging ML forecasting models can also increase effectiveness by reducing the time people normally spend on generating and verifying their forecast.
    2

    How do you prove that an ML-based forecast is better than the existing one?

    During the so called Test stage (model quality and stability testing), followed by the Train stage (model building), we create an ML forecast “in the past” and compare it with historical sales in a selected period, as well as with an existing statistical and/or experts’ forecast which was generated and locked for editing in the past. Then after launch, we track accuracy by comparing actual sales data with the ML forecast, created several weeks ago (usually 4-6 weeks back).
    3

    Which are the key requirements for data to build and run an ML model?

    No less than 2 years of historical data, several other sources of input data that are likely to impact sales (in addition to historical sales which we forecast with ML, and data like promotions, inventory, price changes would be helpful). It is also important that key data inputs have their own future forecast (e.g. promo plan for the next few weeks), which will help the ML model generate a more accurate forecast.
    4

    What do you mean by “demand forecasting”?

    We are experts in building a machine learning-enabled forecast in the area of demand planning:
    -Primary sales forecast (shipments)
    -Baseline sales forecast
    -Promo sales forecast
    -Secondary sales forecast (off shelf)
    -Promo investments forecast and optimization
    -New products sales forecast
    -Optimal shelf price forecast
    5

    What is a forecast horizon in an ML model?

    The forecast horizon depends on the requirements of the customer, and the quality of the forecast within the required forecast horizon depends on the quality of the incoming data and the availability of a forecast for the future, for example, a promo plan for the future. If the promotional plan is available for the period of 3 months in advance, the model will build the most reliable sales forecast for exactly three months in advance.
    6

    Can a model be based only on a single data input apart from historical sales?

    Potentially yes, but then multivariate algorithms will not be leveraged to the fullest. In order to fine-tune the model, it is necessary to have historical sales data, which we will then forecast, and data for at least one key factor, for example, data on promotions (periods, cost of promotions).
    7

    Which business teams will benefit from the solution?

    Though this is defined by the organizational structure, generally we can split teams by business functions depending on a forecast type:
    -Teams in charge of demand planning and sales forecasting: Demand planning in Supply chain and Sales departments
    -Teams in charge of promo sales forecasting and promo investments management: Trade and customer marketing, Sales Finance
    -Teams in charge of pricing: Finance, Business intelligence, Trade and customer marketing
    8

    Which IT tool do you use to create and maintain an ML model?

    An ML model is not linked to any IT tool because it’s a code which generates a file with a forecast in a format set by end-users (Excel, CSV, etc.) When developing and rolling out a model, the code is run on Augmento’s machines. Later on we will be able to integrate the model  into your existing local infrastructure or help choose a relevant cloud-based ML platform.
    9

    Will we need to roll-out an IT planning solution in order to manage our newly launched ML forecast?

    Implementing an ML-enabled forecasting model does not necessarily entail launching any planning IT solution. However, it’s optimal when a planning platform like Anaplan already exists in a business, but that does not define the project’s success. An ML forecast can be exported into an Excel file, into your existing planning tool, or directly into a BI platform for visualization.
    10

    Does an ML model have any limitations?

    The quality of input data and sufficiency of historical data (no less than 2 years), as well as a future forecast for all chosen data inputs, are instrumental in making the model run well. For 2-3 months after launch, the workload for planners increases because they work with two versions of the forecast – their current statistical and a new ML one, in order to then fully move to the ML forecast.
    11

    What do you mean by Manage by exception?

    The forecast generated by the model is split into several segments based on agreed criteria:
    –No touch – this is a forecast (e.g. a sales forecast), which does not require review by forecast specialists and is handed over directly to supply planning/production or sales teams depending on the forecast type
    –Light touch – a forecast which is recommended to be double-checked by experts
    Heavy touch – a forecast which has to be created manually. For example, when it comes to a sales forecast, there are products which have no historical data like innovations and special promo packs.
    12

    How long does it take to launch the model in full?

    It depends on the complexity of the brief (e.g. a model for the whole business or a part of it), data accessibility, etc. On average, the project takes 4-6 months.
    13

    What is specific in your approach to one of the big consulting firms like Accenture, SAAS and others have?

    First, we deliver an open model, not a black box. This means that after launch, a client can develop and support the model themselves without our involvement. Second, our team of data scientists and project managers has extensive experience in working with planning departments of big multinational companies. Third, a relatively low cost of model development and support.
    14

    What involvement from the business is required?

    Mainly, a client needs to participate in helping define objectives for the project, review all the data available, provide needed access to databases and lead integration during the go-live stage and finally do change management for existing processes, ways of working and structures.
    Заказать звонок

    ×