Demand Forecast
Tiki is an e-marketplace with its Retail business operated by Tiki itself. Demand forecasting is a crucial part of any retail supply chain. We built a decision-support system from scratch to automate parts of this process and help our users—Tiki’s internal employees—make more informed decisions.
Summary
Challenge
Manual demand forecasting no longer worked well, falling short in accuracy and coverage. A more reliable and effective solution was needed to improve decision-making.
What I did
Conducted a user observation to analyze the user's current method.
Formed the product concept with PM, DA & Dev.
Designed the UI.
result
The research insights informed the developing & design process.
After launching the first version, overestimation rate dropped significantly to only around 10%.
what i've learnt
Designing decision-support systems effectively requires a thorough understanding of user's decision making process (typically experts in the task), plus domain knowledge, plus technical expertise.
Context
Inventory is a crucial investment in retail. Overstocking ties up cash, while understocking leads to missed revenue. The right inventory level meets customer demand, a challenge assigned to the Procurement team. Each month, they must forecast demand and plan stock procurement and replenishment.
The manual approach worked in the early days when there were fewer products and cost optimization wasn’t a major concern. However, as the business scaled to 100,000+ SKUs, the Procurement team faced stricter coverage and accuracy requirements. The manual process became inefficient, making
automation solutionessential.
Research
To automate and optimize this decision-making process, we need to
study and analyzethe current workflow to identify pain points.I conducted a
Contextual Inquirywith 12 users, each one in charge of different product type (Mom and baby products, Health & beauty products, Electronic devices, …). This method involved observing their end-to-end workflow and interviewing (semi-structured) them during the process.Questions to be answered
How is the demand forecasting task currently performed? What the process is like? What calculation they make?
Why they do what they do?
Identify the problems or challenges which could lead to low coverage and accuracy.
Findings
After collecting the data, my thematic analysis revealed their methods & process, as well as the potential risks it could cause to the task’s accuracy & coverage. The findings, as well as my recommendations for next steps, are put together in the below report.
The report guided our team (PM, Data Analyst, Developers) in further discussions to define the product concept.
Key findings drawn from the research
1. Patterns in the user's decision making
Although different product types (e.g., food versus electronic devices) may follow different processes due to their unique characteristics, there are key similarities that could enable a unified decision-making process. For example, at some steps, they all:
Rely on similar data references (e.g., historical sales data)
Share common goals (e.g., identifying historical sales trends),
And use comparable methods to achieve these goals (e.g., calculate average sales per day).
As I was able to synthesize all of those patterns, these serve as a knowledge base for the data/dev team to build their forecasting model/algorithm/set of rules for our AI product.
2. Ad-hoc factors
During their forecasting process, there are certain ad-hoc or unstable factors require users to rely on intuition and market experience rather than calculations. For example:
Seasonal events (e.g., Christmas, back-to-school) → unusually high demand.
Trendy or outdated products (e.g., driven by TikTok/social media) → unpredictable spikes or drops in demand.
New products → no historical data to reference.
Unusual situations (e.g., supply crises, material shortages, rising gas prices) → demand may exist, but procurement isn’t possible.
These are areas where the system struggles to automate due to limited data and predefined rules, making user input, audit, and review essential.
 |  |
|---|---|
Back to school season products | An example of on-trend product |
3. Behaviors, errors and biases
During the forecasting process, users must navigate multiple Excel files and tools spread across different sources, leading to wasted time switching between platforms and performing repetitive tasks. A significant portion of their time is spent collecting and centralizing data into a single Excel file, followed by manual calculations to derive necessary metrics.
Additionally, human errors and biases frequently arise, such as mistakes from copying and pasting or irrational decision (e.g., relying on overly short time frames for analysis). These issues compromise the accuracy of the forecasts.
These are risks that our product could prevent. Automating these steps would streamline the process, reduce errors, and enhance the reliability of forecasting outcomes.
Design the product concept
The research findings guided our team (PM, Data Analyst, Developers) in further discussions to define the product concept. For the first phase of the project, we decided to build a decision support system with core functions to:
Generates predictions using simple forecasting methods*.
Allow users to review and refine these suggestions using reference data shown in the UI, along with their own market intuition.
We also identified supporting features to improve usability and forecast coverage, such as Advanced filtering, Data sorting, Marking products as reviewed or not, Reason for too small/too large numbers, etc.
Product goals and success metrics were also defined at this stage.
*Our first version generated predictions using simple time series method along with a set of rule defined based on the research findings. More advanced Machine Learning methods were introduced in later stages. Read more at: engineering.tiki.vn
From concept to UI
One of our biggest challenges when we start designing the product was determining which data to display and how—whether as numbers, tables, or charts—to optimize clarity and usability.
Prioritizing which data mattered most for decision-making, along with creating a clear visual hierarchy for them, was something I spent a lot of time on.
Given how data-heavy and complex the product was, my goal as a designer was to create an interface that felt comprehensive yet simple—empowering users without overwhelming them.
While I can’t cover every design decision here, below are a few key examples.
Launch
As we launch our first version,
different metrics were trackedto measure the product success, such as user's review & edit behaviors, product coverage, system's forecast and user's forecast accuracy (compared retrospectively to actual demand).Forecast accuracy improved gradually over time. After 2 months,
overestimation reducedfrom a large margin to just around 10%.In terms of
product adoption, the transition from Excel to our tool was pretty smooth. We reached 80% adoption across 13 category teams after 1 month and by the second month, the demand forecast process was completely transferred to our tool. Despite the unfamiliar interface and task complexity, the learning curve was low—only a few of the 30 users needed assistance.