Intern(Chow Jun Hoe)

 29/8/2025 Today, I focused on putting together a PowerPoint presentation summarizing all the models and data processing methods I used earlier, including Holt-Winters, Decomposition, ARIMA/SARIMA, XGBoost, and the Python/Excel data filtering techniques I learned during my internship. The slides are clearly structured and will be easy to share with my manager and team later.

 28/8/2025 Today, I focused on summarizing the XGBoost model, including feature selection, parameter adjustment, and performance analysis for holiday sales forecasting. I also organized some Python data filtering techniques I learned during my internship, such as conditional filtering, multi-condition combination filtering, and data cleaning methods, to facilitate rapid application in various data processing scenarios.

 27/8/2025 Today, I primarily completed a summary of Decomposition and ARIMA/SARIMA models, covering model principles, application scenarios, parameter settings, and analysis of their performance and accuracy on Eco-Shop data. This summary will serve as part of a review of the overall forecasting methodology and will be used for subsequent reporting and model optimization.

26/8/2025 My main task today was to complete a summary of the Holt-Winters model. In the process, I focused on reviewing the model's three main components: the calculation methods for trend, seasonality, and level. I also compared the applicability of additive and multiplicative models to different sales data characteristics. I also summarized the model's performance in predicting EcoShop sales, including its MAPE accuracy, its adaptability to holiday fluctuations, and its strengths and weaknesses in short-term forecasting and long-term trend analysis.

 25/8/2025 Starting today, my manager has asked me to systematically summarize my learning and work achievements over the past six months. My main task today is to complete a sales quantity forecast based on sales participation, specifically, to predict sales volume based on sales participation. This task requires combining historical sales data with participation metrics, building a forecasting model, and verifying the accuracy of the forecast results.

 22/8/2025 Today, I held a meeting with my manager and executive team to discuss the current XGBoost model settings and prediction results. During the meeting, my manager raised questions about the model's settings, particularly regarding the rationality of the data processing logic and feature selection. I addressed these issues. He also pointed out that some of the current results may be affected by immature data processing, resulting in a lack of confidence in the model's predictions. The meeting ultimately determined that further adjustments to the model are needed, including optimizing the data preprocessing steps,

 21/8/2025 After completing preliminary testing of the TCN model, today I turned my focus to experimenting with the N-BEATS (Neural Basis Expansion Analysis for Time Series) model. Based on the specified data features and the factor of past holidays, I began building the N-BEATS model locally. However, during training and prediction, my computer encountered performance or compatibility issues, preventing the model from producing prediction results.

 20/8/2025 Today, I successfully built a TCN (Temporal Convolutional Network) model. Theoretically, TCN can better capture long-term dependencies in time series, making it particularly well-suited for data with cyclical patterns and seasonal fluctuations. However, after applying the model to EcoShop's sales data, the results showed that the TCN was poorly suited to the data, failing to outperform the existing XGBoost model in terms of prediction performance.

 19/8/2025 After completing multiple rounds of optimization for the XGBoost model, I shifted my research focus to deep learning methods. I then began systematically learning how to use the TensorFlow framework and began building the infrastructure for the TCN (Temporal Convolutional Network) model.

 18/8/2025 Today, I continued optimizing time series features based on XGBoost, focusing on adjusting the number of lag days to observe their impact on forecast performance and MAPE. I also began researching the Temporal Convolutional Network (TCN) model and experimented with applying it to sales forecasting.

 15/8/2025 Building on the optimization of the rolling forecast model, I added cosine and sinine transformations based on periodicity to handle date information in time series. Verification results show that this method maintains forecast consistency across year-ends (especially for dates like 1/1/2025), effectively avoiding forecast instability caused by date code jumps.

 14/8/2025 Today, I successfully converted my XGBoost model to a rolling forecast mode, achieving a significant reduction in the MAPE. This demonstrates that this step-by-step forecasting approach is indeed helpful in improving overall forecast accuracy. However, the rolling forecast mode fails to effectively account for some holiday factors, particularly sales fluctuations near holiday times. Further research is needed to identify solutions.

 13/8/2025 Today's main task is to convert the XGBoost model to a rolling forecast mode, so that the step-by-step forecasting process is more closely aligned with real-world business needs. This adjustment not only involves modifying the structure of the forecasting process but also replanning the split of training and test data to ensure that each forecast step is updated based on known historical data.

 12/8/2025 In the morning, I mainly assisted with the transportation of supplies for the charity event, loading the relevant goods onto the truck in an orderly manner. In the afternoon, I focused on optimizing the XGBoost sales forecasting model. This included checking feature importance, adjusting parameter combinations, and experimenting with introducing new data features to improve the accuracy and stability of the forecast.

 11/8/2025 Today’s main task is to continue adjusting Coded OOS, and after completion, continue to study XGBoost model optimization.

 8/8/2025 My primary task today was to present the previously simplified Coded Out-of-Stock (OOS) information to the relevant teams and engage in in-depth discussions with them to evaluate whether the simplified version meets business needs and actual application scenarios. During the discussion, everyone provided feedback and raised questions regarding the feasibility, accuracy, and compatibility of the simplified logic with existing processes. After analysis and discussion, I discovered that the current simplified solution, in some cases, failed to fully reflect the true cause of the out-of-stock situation, potentially leading to biased subsequent data analysis and decision-making. Therefore, we ultimately decided to restructure and restructure the version to ensure it remains concise while still meeting the business department's usage standards and accuracy requirements.

 7/8/2025 While analyzing historical sales data today, I discovered an interesting and important pattern: sales often surge before a price increase. This phenomenon is likely due to consumers stocking up in advance upon learning or sensing an impending price increase, leading to a short-term surge in sales. Based on this insight, I added price-starting price-related features to my XGBoost model, such as flagging a special variable for a few days before the price-starting date or introducing a dummy time window to reflect this tendency for early demand release. By incorporating this feature, the model can better capture the impact of price changes on consumer behavior, thus reflecting the sales increase before the price-starting price in its predictions.

 6/8/2025 My primary focus today remained on tuning the XGBoost model. By further optimizing the input features, I successfully found a way to enable the model to consistently capture the sales decline after the Raya holiday. This discovery is of practical significance, as Raya is one of Ecoshop's most critical sales periods of the year, and sales typically experience a significant decline after the holiday. The model's ability to identify and reflect this trend will help improve overall forecast accuracy and business response capabilities.

 5/8/2025 Today's work progressed relatively smoothly, primarily focusing on tuning and optimizing the XGBoost model. The focus was on fine-tuning model parameters and re-examining input features, particularly after incorporating external variables like holidays, starting prices, and seasonality to observe how the model responds to sales trends.

 4/8/2025 My main task today was to simplify and organize the "Coded OOS" information, as directed by the Executive. The original Coded OOS data contained numerous technical tags and fields, making it difficult for non-technical personnel to read and understand. This task aimed to logically categorize, consolidate, and semantically simplify the original 29 Coded OOS flags, making them easier to understand and present.

 1/8/2025 This morning, I mainly completed the unfinished packaging of the charity supplies from the previous day, ensuring that all Eco-Shop-sponsored items were correctly categorized and packaged. In the afternoon, I returned to optimizing the XGBoost model, further adjusting parameters based on the results of the previous few days' testing. For example, I adjusted the lag time, such as the number of days.