As the year draws to a close, and another one starts we are reflecting upon our yearly risk forecast performance  for 2023.

How we gather, process data to build trend forecasting models

FOODAKAI has over 200 hundred millions of records that can address a range of questions for food safety risk. FOODAKAI’’s data is processed by our AI powered data processing algorithms and curated by our food safety experts. Our models provide forecasts for the incidents trend using historical data of over 20 years. We are building different trend forecasting models for each food category using historical food safety incidents (food recalls and border rejections). To train and test our models we are utilizing only high quality and accurate data that are reviewed and validated by our domain experts. An accuracy metric is used to measure the performance of our models and the validation of the trend forecasting models is done using a cross validation approach.

Our 2023 Forecasts Trend: a summary

In our Key Food Safety Trends 2023 Report, we presented the incidents trend and emerging risks for the top 10 most important food categories, as they were selected by food safety experts. For most of the food categories a decreasing trend of incidents was forecasted (table1). The largest increasing trend was forecasted for Poultry & Poultry meat products and the largest decreasing trend was forecasted for Cocoa, Coffee and Tea products.

As we enter 2024, it's now a good time to see how our forecasting models have performed.

Trend forecasting performance analysis

Using the actual reported incidents by Authorities in 2023 we conducted a performance analysis for our forecasting models which is presented in table 2. The performance analysis was conducted by comparing the forecasted trend with the actual incidents reported by the authorities. Incidents include both food recalls and border rejections (import refusals). The correctly forecasted trend and risks are highlighted in green. The table shows that for 6 out of 10 food categories, FOODAKAI’s models correctly forecasted the trend of incidents. In addition to the same can be said the majority of the increasing and emerging risks which were also forecasted successfully.

Let's see the performance analysis results for each food category.

Dairy products

Our models forecasted correctly that we will have a decrease of incidents for dairy products and indeed we had less incidents in 2023 compared to 2022. As for the increasing risks, our models forecasted correctly that the risk of listeria monocytogenes in dairy products will increase in 2023. For emerging risks our models forecasted that Cronobacter Sakazakii will be an emerging issue following 2022, where many incidents involving the hazard were recorded. However, the trend of Cronobacter risk in dairy products decreased and a new risk for Yersinia Enterocolitica was reported.

Cereals & Bakery products

For cereals the forecasted trend of incidents was very close to the actual trend of incidents reported in 2023. More specifically, the actual trend of reported incidents was -24,5% compared to the forecasted trend which was -22%. As for the increasing and emerging risks, our models forecasted correctly that the risk of pesticides and specifically of ETO would increase in cereals and bakery products. In addition to that, PAS was correctly forecasted as the emerging risk in cereals and bakery products.

Herbs and Spices

For Herbs and Spices our models forecasted that there will be a decrease of incidents and indeed the actual incidents decreased compared to 2022. The percentage of the decrease was higher (11%) compared to the one that was forecasted (6,5%). Regarding the increasing risks, our models correctly forecasted that the risk of pesticides and ETO specifically will increase. For the emerging risks, the forecasting models correctly highlighted several pesticides but did not manage to identify at the beginning of 2023 the mineral oils as a new risk for herbs and spices.

Fats & oils

The trend forecasting models that we have developed for fats and oils did not manage to successfully forecast the trend of incidents nor the increasing and emerging risks. More specifically, our models forecasted that there will be an increase of incidents by 10.8% whereas the actual incidents were decreased by -14%. The low performance of the model may be due to a small amount of historical data that we have for fats and oils. This category was also very much affected by the sociopolitical instability of regions such as Ukraine which are very important producers of oils.

Cocoa, Coffee & Tea

Our AI models forecasted with high accuracy both the trend of incidents but also the increasing and emerging risks. More specifically, a significant decrease of incidents were forecasted (29,75%) and indeed the decrease of the reported incidents as of 10/12/2023 very close to the forecasted one (29%). It was forecasted that the risk of pesticides like ETO and heavy metals will increase, and indeed the risks for Cadmium and pesticides increased. Piperonyl butoxide was successfully highlighted at the beginning of 2023 as an emerging pesticide for coca, coffee and tea products. The high performance of this forecasting model may be due to the availability of a high amount of data and the seasonality that we have in the historical data.

Nuts & Seeds

Our trend forecasting models did not manage to correctly forecast the trend for nuts and seeds. Based on the actual reported incidents for nuts & seeds we have a small increase of the number of incidents (4%), whereas the models forecasted at the beginning of this year that there will be a significant decrease of incidents following a previous pattern of having less incidents after a peak of incidents in the previous year. However, the AI models still managed to correctly forecast the increase of risk for pesticides and Mycotoxins. As for the emerging risks, the models correctly forecasted that in 2023 we will have some emerging pesticides risks but did not forecast the emerging risks of new Salmonella spp. serotypes.

Fruits and vegetables

For fruits and vegetables our models forecasted that there will be a small increase in the number of reported incidents (4%) but the actual number of incidents decreased by 6.5%. However, our AI models forecasted correctly the increasing and emerging risks. More specifically, the increase of the Listeria monocytogenes, Salmonella and Aflatoxin was correctly forecasted. As for the emerging risks, mineral oils were correctly forecasted but this was not the case for E-coli and Salmonella spp..

Meat & Meat Products

Our AI models correctly forecasted that there will be an increase of incidents for 2023 for meat and meat products other than poultry. However, the actual increasing trend (46,4%) was much higher than the one forecasted (7%). The increasing risks for Listeria monocytogenes, Salmonella spp. and fraud issues were successfully forecasted by our AI models. In addition to that the emerging risk of Pyrrolizidine alkaloids (pas) was successfully forecasted.

Poultry & Poultry Meat Products

For poultry & poultry meat products the forecasted trend (11,8%) for the reported incidents was very close to the actual trend (10%). As regards the increasing risks, the AI models forecasted successfully that the Salmonella spp. and foreign bodies risk will increase. The models successfully forecasted that we will have several new salmonella serotypes as emerging risks in poultry meat.

Confectionary products

Following many issues due to biological hazards in chocolate products, our models predicted that there would still be a significant increase of incidents reported by the Authorities. This was not validated by the actual food safety incidents published by the authorities.

We can see that even when the trend of the incidents is not forecasted correctly, the emerging and increasing risks are predicted correctly. This is the case with fruits and vegetables and confectionery products. This is highly important as the forecasting of trends of specific risks’ can be extremely useful for the early adjustment and reconsideration of preventive measures.

Risk forecast for geographies

The performance analysis for the regions with high risk profile are presented in table 3. As can be seen for most food categories (8 out of 10), our forecasting models highlighted correctly at least one of the regions with a high-risk profile.

What this validation tells us about trusting such forecasts

According to the performance analysis, the forecast models can correctly forecast the incidents trend, the emerging risks for the food categories and the regions with higher risk with an accuracy of more that 60%.

For categories for which we do not have a lot of data for historical incidents, such as Fats & Oils, the accuracy is lower compared to categories for which we have a lot of data. The low availability of the data is an important issue that we need to keep in mind when using such forecasting models.

Which are the limitations of forecasting models in food safety

When using forecasting models in food safety we need to keep in mind the following limitations

• We cannot have accurate models if we do not have enough historical data. The historical data needs to be stationary timeseries, this means to have seasonality and trends.
• The AI models cannot forecast unknown risks for which there are no previous knowledge e.g. no reports announced by Authorities, research studies and reports.
• It’s very difficult to predict how socio-economic situation in specific regions (e.g. GDP, corruption index) might affect risk in foods.
• It is not possible today to forecast events that depend on many parameters that cannot be predicted.
• We cannot forecast events such as the behavior of people on how they handle an ingredient in the production line which may introduce a risk in the supply chain.

Which preventive measures such forecasting models can support.

The forecasting models can be a powerful tool in order to better design and outline preventive measures that can reduce the likelihood of having an important issue in your supply chain. More specifically, you can formulate a more focused lab test plan by including the hazards for which you will have increasing risks. In addition to that, you can adjust the audits plan by focusing on suppliers that you source ingredients from which are more susceptible to risks. The actions that you can enhance and support when using such risk forecasts include:

• Communicating the risk to organization
• Proactively adjusting supplier practices
• Proactively adjusting your testing plan
• Proactively adjusting your audit plan
• Changing suppliers

Expert Reflections on the Review

"Continued refinement of the AI food industry models, will help further minimize food safety issues and enable food safety leaders to use the FOODAKAI crystal ball to see around corners"

Neil Marshall, Managing Director at Guv Consulting International LLC

Want to see how you can benefit from such forecasts? The arrange a call here.