Introduction

To manage our equipment stock levels, including equipment availability for exports, it is necessary to forecast the impact of weekly operations in terms of the expected import returns (stock increase) and export pick-ups (stock decrease). To ensure that a forecast is always available to any active site, ROCK relies on the product "Auto Forecast" from Maersk Digital.

Features and Benefits

Benefit	Description
Complete picture of global supply/demand	As some locations are not adding forecasts manually, the automated forecast enables a a complete and holistic overview of global equipment supply/demand.
Data-driven approach combined with local knowledge	As a principle, the auto forecast can always be refined by local equipment planning teams, by manually overriding the auto forecast values. As such, the auto forecast serves the local forecasting teams as an additional data baseline or as an extra input for analyzing and possibly adjusting their forecasts. The local forecasts are referred to as the "Manual Forecast". If no manual forecast values are provided, ROCK will use the values from the auto forecast to determine the effective forecast value in the stock management screens. Managing the forecast is done in the dedicated forecast screen in ROCK (see Manage Forecast). Furthermore, to guide equipment planners to make accurate forecasts, ROCK also includes tools to evaluate the historical forecasting errors, for manual and auto forecast respectively. This can be done in dedicated screens (see Forecast Accuracy Report).
Granular predictions	The forecast is produced on Pool level, for equipment groups and per operator (e.g. NLROT/40HIGH*/MSK). This granularity allows a detailed forecast which further enables equipment planners to make reasonable adjustments at the right level.
Correlations	The auto forecast is build to consider past trends and seasonality, currency exchange rates, and national holidays.
Quantified accuracy of predictions	Using standard deviations from the forecast, probability theory allows us to quantify accuracy of the forecast prediction. The certainty can be translated into the necessary size of buffers required (see Target Stock Levels v1). This calculated approach allows target stock levels to be set automatically.
Business Rules	As part of the ROCK project, a set of business rules (see RKEM Mapping) was defined in a separate workstream, together with "Agile Data Labs" team and business SMEs. This enables an accurate forecasting baseline for the future predictions.
Forecasting Correlations	The accuracy of a prediction can (often) be improved by correlation of certain variables. For example the number of ice creams sold (variable x) would likely correlate with the temperature (variable y). As such we could assume there is a correlation relationship between the two variables. The same approach is introduced in the auto forecast to increase accuracy of the predicted number of export pick-ups and import returns. The list of forecasting variables introduced in the forecast includes: Seasonality One of the benefits of having large sets of data (5 years used for auto forecast) is that it is easier to identify seasonal trends, and the relative strength of the trend. Seasonal trends are identified and applied to future prediction of export and import events. Holidays By using a data set of known holidays per country, patterns of volume development of pre-holiday, holiday and after-holiday week are identified. Based on these correlations and known future holidays for corresponding weeks we adjust future forecasts. This has proved to give a significant accuracy improvement in predicting future customers’ behavior. Currency Exchange Rates Exchange rates of different currency pairs are often affecting an increase or decrease in sales volumes for our customers. In case rates are favorable for buyers – orders to manufacturers will increase and vice versa. On export side it might often give us around one month from the known currency exchange rate change triggering orders by buyers to be placed, goods manufactured and packed, transportation booked etc. and until the empty container is picked up for export. On the import side, our advantage is even greater as after exchange rate triggers a change in exports and all the above-mentioned events are finishing we need to add transit time until container reaches the destination and is returned empty – often around two month from the actual exchange rate change.

Building the Auto Forecast

Creating a Forecasting Baseline

The forecast is produced from a baseline of 5 years of past actual import return and export pick-up events, using the rules the RKEM Mapping. The baseline is prepared by BI team, outside of ROCK.

Geo data

To ensure reasonable accuracy of the forecast data, all historic import and export data is aggregated from site (depot, terminal) data to pool geo-level (all sites in a pool, e.g. Rotterdam), using latest GEO relations (see Geo Data). With this approach, all history at site level will contribute to the currently related pool, thus disregarding past the changes in geo relations over time. Geo relations are managed in the SMDS application (Single Master Data Store).

For example, if a major depot AAXXX01 was part of the Pool AABBB from 2010 to 2013 and then from 2013 and onwards up to now it was linked to pool AACCC, then we will assign AAXXX01 volumes to AACCC from 2010 till now, as per latest GEO relations.

Equipment data

To ensure a reasonable accuracy for the forecast, each forecast dimension is limited to equipment types on group level (see also Equipment Types). Currently the forecast is only produced for equipment groups that represent a reasonable part of the container fleet. An overview of the equipment groups included in the auto forecast is described in the table below.

DRY

REEF

SPEC

Included

20DRY*

40DRY*

40HIGH*

45HIGH*

20REEF*

40HCRF*

40CASC*

40CARF*

20OPEN*

40OPEN*

40FLAT*

Not included

40MSFU*

40AFRF*

40XFRF*

20FLAT*

40TWDK*

Retraining the Model

The forecast is generated for each pool available in the geo relations. As export and import behavior differs from pool to pool, the model considers this by selecting one of 15 forecasting models that perform best to the specific pool. The approach sets a forecasting baseline of a simple forecasting model 4 weeks rolling average and calculates the mean error. Using 15 different mathematical forecasting models (including forecasting correlations), each model will try to outperform the mean error score of the 4 weeks average. The model with the lowest mean error score is applied when the forecast model is run next time, which means that the 4 weeks rolling average will be used as a fall-back model.

The model is retrained weekly, to consider latest geo relations between site and pool in the retraining of the model. This effectively means that a suitable mathematical model is identified and set for each pool/equipment group/operator combination.

Generation the forecast

Based on the retraining, the appropriate forecasting models are selected to each dimension, and predictions are created. As some locations does not include enough data to produce an operator forecast with reasonable accuracy, the auto forecast engine will merge all operators and forecast the pool to a single number. In this case, ROCK will receive the data on a merged operator code "ALL" and will split the forecast value across operators using relevant split ratios.

When forecast is received in ROCK, the predictions are split from pool to site, using relevant import/export split ratios, which is necessary to support granular equipment management (see Split Ratios).

Improvement Ideas

As the auto forecast is directly used to automatically calculate buffer for stock level (see Target Stock Levels v1), an increase in forecast accuracy can directly be translated into less required buffer globally. Consequently, a higher accuracy means a smaller container fleet required, and thus higher container utilization achieved.

Future developments to increase forecast accuracy includes

Forecastes from Future Pricing Optimisation, FPO (formerly known as Revenue Optimization programme)
GRIs,
Market Share Forecasts
Commodity Indexes
World Economic Outlook (WEO) from IMF (volume of Exports and Imports of goods per country)
Using Imbalance (Import-Export) as correlation measure

ROCK Documentation : Auto Forecast