meantime may you share link to another data set to download for demand prediction ML task? As you wrote : Why you should never forecast sales On the other hand, sales are constrained by (lack of) supply. So data set to predict demand but for real live situation when sales are constrained by (lack of) supply
www.datasource.ai/en/home/data-science-competitions-for-startups/vn1-forecasting-accuracy-challenge-phase-1/description Sure, there you go. We had to move the starting date to 12th of September
Hello, if you want to learn how to make your own models: www.amazon.com/Data-Science-Supply-Chain-Forecasting/dp/3110671107 (It's a step by step approach starting from 0, so don't worry if you are not an expert today.) If you want to understand how ML impacts demand planning and how your teams should work with it. www.amazon.com/Demand-Forecasting-Practices-Nicolas-Vandeput/dp/1633438090
Hi I am an S&OP manager based in Pakistan, I wanted to know how can I get access to your book? also I have been working to create baseline forecast in my organization but since there are no historic demand driver details available its very difficult to generate baseline forecast, any mathematical approach that I can use to atleast begin with forecasting for longer period of months?
Hi Nicholas, when it comes to feature engineering for future covariates, which features are a must according to you? The only future feature I've been able to implement is the lag features, however one is then constrained by the lowest lag feature, i.e if you have lag 7 day feature, you can't predict further than 7 days into the future. What other future variables are there that one would know in the future, apart from holidays and company specific things like marketing costs, promotions etc?
🎯 Key points for quick navigation: 00:29 *📊 Forecast Value Added (FVA) assesses how different teams contribute to improving or worsening forecasting accuracy.* 02:05 *🔄 Demand planning processes typically involve automated baseline forecasts adjusted by teams to enhance accuracy.* 04:19 *🎯 FVA aims to ensure forecast accuracy improvements without excessive time spent on adjustments.* 05:02 *📉 FVA framework tracks how each team's adjustments impact forecast accuracy positively or negatively.* 08:16 *📈 Comparing forecasts to benchmarks like moving averages helps assess the added value of forecasting models.* 11:16 *🎯 Setting accuracy improvement targets relative to baseline performance can be more effective than absolute accuracy targets.* 14:41 *💰 Evaluating forecast errors based on value helps prioritize improvements on high-value products over low-value ones.* 19:28 *🌐 Forecasting across various time horizons (short, medium, long-term) supports strategic supply chain decisions.* 23:09 *📊 Forecast Value Added (FVA) helps identify SKU-level performance, guiding decisions on where to focus and where improvements areneeded.* 23:38 *🔄 FVA encourages a positive feedback loop by comparing market performance against statistical baselines, fostering model improvements.* 24:31 *🌐 Different forecast horizons (short-term vs. mid-to-long-term) require varying model strengths, prompting discussions on model integration.* 25:12 *🤝 Collaborative discussions using FVA help align marketing and finance teams by highlighting where judgmental adjustments add value.* 25:49 *📉 Separating positive and negative adjustments in FVA reveals insights into which adjustments enhance or diminish forecast accuracy.* 27:01 *🎯 Forecasting supports supply chain decisions, aiding in manufacturing and procurement planning crucial for business operations.* 46:55 *🌍 Different countries and industries may require tailored risk management strategies in pharmaceutical production to ensure patient needs are met without compromise.* 47:22 *🤝 Collaborative relationships between planning teams and sales are crucial for mitigating forecast overrides, emphasizing education on supply chain dynamics and outcomes.* 48:27 *📊 Presenting a range of forecast possibilities enhances decision-making by providing stakeholders with more nuanced insights and flexibility.* 49:20 *💡 Implementing statistical engines requires effective change management strategies to shift from manual to automated forecasting processes, emphasizing education and gradual adoption.* 51:12 *💼 For small to medium-sized businesses, affordability and implementation time of forecasting tools can pose significant challenges despite their potential benefits.* 54:04 *📈 Transitioning from manual to automated forecasting involves proving benefits through accuracy metrics and building confidence in system outputs to foster acceptance among demand planners.* Made with HARPA AI
Thanks for the vides Nicolas! I have read all your books - what a fresh take! Please make a video for niche demand planners and inventory controllers to which I belong. That is MRO spare parts inventory which is fraught with intermittent demand and skewed probability distributions. If you ever revise your books 'Data Science for Supply Chain Forecast' and 'Inventory Optimization' please include these topics. I read your article on Croston Method in towards data science and it was very well written with a practitioner's perspective. Whenever possible please make a video on demand forecasting and inventory control for spare parts. Thanks!
Love it! 🙏 Thanks for sharing Nicolas! So once we have a great forecast using these three steps, how does one implement the inventory optimization element? I’d be curious to hear your top 3 on the IO portion. Of course, your IO book goes into this quite well already!
Hi Nicolas, it is a very explanatory webinar about outliers. I have a question. Do we need to apply outlier detection process based on train data, or whole data (train+test)? I hope my question is clear.
I would try not to do any statistical outlier detection. I would invest more in data cleaning. If you remove outliers from the test set, you are somehow overfitting - so I would not do it.
Hello dear , I want to congratulate you for the content and the skills that you teach people . I have a question : the forcast ,like you describe it , is applicable for the retail demand planning and the MTO strategy , that's right .Because I don't think that this kind of forcasting is relevent in an MTO industry where the demand is not stable .
Hello, you can apply it to MTO but it might require some differences. For examples you could focus more on forecasting raw material, or include as a features in your ML engine preorders. Or contractual terms/budget.
Thank you for your webinar. I have a question regarding outliers. I am conducting a serum biomarker research (medical research) consisting 50 patients vs. 50 controls. I have 3 cases having non-detectable values (above the detection level) in the same group. This group is already have higher levels than the other one. I do not want to remove those cases and lose the data. Which strategy should I use ? Should I imputate them with the mean value of the relevant group? or Should I enter the measured highest value/s instead of undetectable ones ? or else ? Thank you in advanced.
Hello, sorry I specialize in supply chain demand planning - I don't think I am legitimate or have the experience required to advise you regarding how to conduct medical research. All the best!
Hi Nicolas, to calculate the bias in case of many unpredictable new products introduction and phase out, the latter are not considered because the time series are not available in the forecast period while the former are considered and demand is greatly overestimated, as the 2 cases do not compensate, the overall bias on the product portfolio is always positive. How do you recommend managing this case? Should I consider as a forecast error also the forecast 0 on NPI even if their time series were not available at the time of the forecast?
Hello, this is quite a complicated case. You could compute bias in two different flavors, with and without NPIs. The idea is that you don't want to bring the message that bias is close to 0% whereas obviously, you missed 10% of NPIs. But the responsibility for these NPIs might lie with another team.
Thank you Nicolas. Actually what happens is that if I include NPIs I get an unbiased forecast overall because NPIs compensate unpredictable phase out products, while if I don’t include NPIs I get a positive bias (globally on the product portfolio). Maybe global metrics in this case are not meaningful and I should look at the distribution of Bias/MAE of product time series.
Awesome session! I'm curious, how would we forecast zeroes? lets say we have inventory for such items but they do no sell at particular time period may be.
Hello Nicolas, what do you think of training a ML model using as input in addition to past demand also the previous month ML forecast enriched by sales? For example, to predict December 2023 demand (M+2) I would use as input features summarizing historical demand + the forecast submitted last month so in september for december (which was M+3) possibly enriched by sales. So if sales enrich a forecast because they are aware of future trends, the following month this information will be captured by the model.
Hi Nicolas, thank you for sharing this. I have a question for you on forecasting error metrics, I know you don’t like MAPE and I agree, but what do you think of WAPE i.e. sum of SKU (actual - forecast) divided by sum of all SKU actuals ? I think it’s a quite good accuracy metric and also easy to explain to business stakeholders as it is a percentage.
BonjourMr , pourquoi vous ne fairiez pas une formation en ligne (payante bien entendu) où vous enseignez le demand planing d'une façon théorique et pratique avec des cas réels , des exercices de prévisions sur excel .....? Nous sommes une génération qui n'aime pas trop recevoir l'information en lisons (même si je ne doute pas que le contenu du livre est pertinant )
Thank you for posting these webinars. Even with all the Q&A on shortages I'm still confused on figuring out unconstrained demand. On your slide you say to bypass it, but in your book it says to censor it. Are you meaning the same thing for both the slide and book? Also, on the slide in this webinar, that's also book, it looks like you're using a default value for demand, which looks to be the last demand value before the shortage for the duration of the shortage. Is that what you use? Your book mentions forecasting techniques that might help estimate unconstrained demand, but I can't find any examples. Can you share those techniques? Do you use machine learning techniques, or an equation? Thanks!
I usually don't use equations to clean shortages. Nowadays, I just censor them. nicolas-vandeput.medium.com/forecasting-demand-despite-shortages-fee899120c08
What great content! I just finished reading demand forecasting best practices and found this video in the footnotes. Very cool, several learnings, thank you!
How do you capture a demand for a manufacturer in a b2b setting. As orders are been placed and stored in the erp system. Do you use the quantity of order placed as the are intermittent in nature.
Track historical orders (and even preorders) and censor periods with shortages: nicolas-vandeput.medium.com/forecasting-demand-despite-shortages-fee899120c08
Please make video on forecasting intermittent time series data. I tried croston, tsb etc but results are pretty bad.I have only 8 months data . Will you please suggest some methods.
I just have a question on the first one; why do we focus til' M5; why not further and then how further do we forecast? Like a dynamic programming problem; we can keep focusing til' the end of the planning horizon to assess what's a good position at M5, M4,... right?
Hello Pras, You have two problems: - On which horizon should you focus your forecasting effort - On which horizon should you focus your planning effort For both, if you use models (anything automated), you could do as much as possible. But if you need human resources (to do the baseline or enrich a model), you'll have to focus on what's the most important. You only have limited time/resources.
Would be interesting to get you opinion on MAPE to compare multiple forecasts (or to use as performance metric for to evaluate multiple time series), since RMSE, MAE are not suitable to do so.
@@nicolasvandeput-SupChains how would you use MAE to compare different product on different scale? Since the MAE does reflect the different scale and is therefore hard to use for comparison or as an aggregated metric of multiple products.
I just noticed your definition of MAE might be different to the standard one (en.wikipedia.org/wiki/Mean_absolute_error) since you represent it as a percentage value. Would be great if you can clarify this.
@@davidtiefenthaler7753 MAE scales perfectly if you have many products. %MAE doesn't scale across different product. it's all explained here: - www.manning.com/books/demand-forecasting-best-practices - towardsdatascience.com/forecast-kpi-rmse-mae-mape-bias-cdc5703d242d In general, no KPIs are perfect. Especially when looking at broad portfolio.
please make a video on forecasting of slow moving intermittent and lumpy demand patterns such as those encountered in MRO parts demands. How to use Croston method to forecast mean demand and its variance/std dev and then how datascience forecasting can help in such cases.
@@nicolasvandeput-SupChains TSB performed better. but this due to the volatility of my demand data. but if one has a relatively stable demand then a 12 months rolling forecast is suitable. But as you know @Nicolas whatever the technique for forecasting you intend to use will really depend on your case study. Thanks Nicolas for responding reading your book currently.
In the next edition please include the forecasting and inventory control models for intermittent items especially spare parts. For example MCROST and (S,S-1) and how to implement them in Python.
I've published an article on intermittent items here: nicolas-vandeput.medium.com/how-to-forecast-intermittent-products-c5d477b90176 You can see how I simulate policies in python here: towardsdatascience.com/make-your-inventory-simulation-in-python-9cb950da8cf3
Thanks for this webinar. You did a great job giving a high level explanation on the ML concept. I was expecting to have comparaisons between models. Data Scientists should focus more on showing the results of experimentation then advertising the concept. I still don’t know of I should invest money and time to build a POC
I discuss results in various case studies. Here are some, Manufacturer with promotions: 20% forecast improvement nicolas-vandeput.medium.com/forecasting-case-study-ml-driven-forecasts-for-a-manufacturer-with-promotions-3a4dea8a9160 Chemical company: 20% forecast improvement nicolas-vandeput.medium.com/forecasting-case-study-with-a-chemical-company-35d02256667e Pharma distributor: 25% forecast improvement nicolas-vandeput.medium.com/an-end-to-end-supply-chain-optimization-case-study-part-1-demand-forecasting-2f071b81a490 Retailer with promotions and pricing: 30% forecast improvement nicolas-vandeput.medium.com/using-machine-learning-to-forecast-sales-for-a-retailer-with-prices-promotions-aab9b35d16a