A Machine Learning Approach to Lead Time Prediction for Custom Components in Yacht Interior Outfitting

In the outfitting of luxury yacht interiors every piece is made to a drawing for a single vessel, and planning relies on processing durations fixed by rule or left to experience. This thesis attempts to estimate with machine learning the finishing lead time, the time a piece spends in painting, from the movements that the production management portal of a firm in the sector records for each piece. The data cover four commissions, three new builds and a refit, and 23 months of entries into painting, from which 14,245 finishing cycles are rebuilt. Eight algorithms are compared with four baselines, simple rules that require no training, over 17 consecutive months, with the models trained only on the cycles already closed at the date of prediction, and the differences are submitted to a statistical test. At entry into painting the chosen model, a random forest, errs by 14.95 days on average, against 15.24 for the family median, the median duration by material family, wood or lacquer, and the difference cannot be told apart from zero. Grouping the pieces before prediction, by domain criteria or with an unsupervised algorithm, does not improve on the single model in an established way. The most solid result is the estimate updated during processing. When the first pieces of the batch a piece travels with return, the updated estimate for the pieces still in painting cuts the error of the family median by 27.9%. For three pieces out of four, in fact, the recorded exit date is that of the batch, and the decision of the department on which batch to process and when does not appear in the data. For a firm that records movements at batch level, the updated estimate and a date to promise computed as the prediction plus a margin, which leaves 34.2% of the pieces late as the family median does, against 66.8% for the rule in use, can already feed a planning tool. Recording the start and the end of the processing of each piece would make observable its duration and the order in which the department processes the batches. The model would then have information specific to each piece and many more independent observations to learn from, with the possibility of predicting better.