Week #5

Week #5 #

Feedback #

Sessions #

During this week team conducted 3 sessions with potential neutral users to get objective observations. The following questions in telegram were asked:

  1. How clear was the interface for you?
  2. How convenient was it to use the service as a whole?
  3. How satisfied are you with the quality of the prediction/analysis?
  4. Is there anything that caused difficulties or remained unclear?
  5. What would you improve in the interface or functionality?

Below are the quotes!!!

Dmitry Ivanov (Experienced investor): #

  1. The interface is clear and quite simple. But there are some things that I would change for convenience.
  2. On a scale from 1 to 10 — by 7. I would move the news above the chart and make them into posts that can be scrolled. This way they will not overlap the chart and will be convenient for familiarization.
  3. I cannot evaluate the prediction yet, because I have not tried to use the system’s signals for real trading. But their display is convenient to watch — that’s a fact.
  4. There was a difficulty switching the types of metals, but then they explained that so far only gold is working.
  5. I would add a risk display and a pull-up of the real quotation chart online.

Tg - @Dimas_PhT

Eugene Marchuk (Engineer): #

  1. Most likely, it is understandable. 🤔
  2. Intuitive. 👍
  3. As if it were far from the truth… 🤔
  4. When you try to click on the “silver” and “zinc” buttons, the charts for gold and the previous statistics/forecast are displayed. 👎
  5. For statistics and forecasts for the week, month, and year, there are “currently unavailable” bars. It might be worth placing the same bars on the silver and zinc buttons. Or fix the graphs for silver and zinc.

Tg - @eugene_marchuk

Vladimir Bazilevich (ML engineer): #

  1. The interface is very simple and intuitive.

  2. I have experience working with QUIK, which is a very clumsy tool in terms of design fashion, but it’s hard to compete with it in terms of data display. Animations on the graph when hovering over a point (especially when looking at several neighboring ones) rather, they interfere.
    To understand the time axis, you have to read the entire label. In the same QUIK, the general part of the label is aggregated — the hour markers and the general date under them. The chart itself is not informative enough from the point of view of trading — there are no opening and closing prices. Candlesticks are more informative than points with interpolation.

  3. It showed some kind of trend, but again, questions for labels on the value axis.

  4. The grid doesn’t contrast enough — you have to strain your eyes to see it. The grid is important when comparing different points in time, especially considering the discreteness of the timestamps.

  6. News can be displayed as pop-up windows on the chart, opened by click or hover.

  7. Meta-information on the news (time of publication, source) would be placed at the beginning of the note.

  8. In general, the eyes get tired of the interface.

  9. Things that are not ready, I think it’s better to hide — they annoy by their absence.

  10. The scrolling of the news is too shallow — it is difficult to notice the buttons themselves, the active area is small.

Tg - @vbazilevich

Analyze #

After analyzing feedback from three independent users, the following key conclusions were drawn:

  1. Interface
    The interface was generally rated as clear and simple. However, users pointed out that:

    • Visual comfort needs improvement — the interface strains the eyes;
    • A more flexible or adaptive UI design is recommended to reduce eye strain;
    • The chart grid lacks contrast, making it difficult to read, especially when comparing points in time;
    • The time axis is hard to interpret — labels are too verbose. Consider aggregated time formatting, like in the QUIK terminal.

  2. Chart and Data Visualization
    The current chart implementation was seen as insufficiently informative for trading purposes:

    • Replace the line/point chart with a candlestick chart to show open, close, high, and low prices;
    • Disable or simplify hover animations, which interfere with the analysis of nearby data points;
    • Improve axis formatting, especially the time axis, to enhance readability.

  3. Metal Type Switching
    Users encountered issues when switching between metals (silver, zinc):

    • Some buttons display incorrect or duplicate data (gold data shown for silver);
    • If data for a metal is not available, clearly show a “not available” placeholder to avoid confusion.

  4. News Section and Contextual Information
    Feedback indicated the need for structural changes in the news display:

    • Move meta-information (publication time and source) to the top of each news item;
    • Consider displaying news as tooltips or pop-ups on the chart tied to relevant events;
    • Place the news section away from the chart, so it does not block important data;
    • Scrollbar and buttons for news are too small — users had difficulty interacting with them.

  5. Prediction Model and Analysis
    Feedbach showed that model need to be updated with the graph itself:

    • Improve prediciton status.

  6. Technical Functionality
    Several technical issues and recommendations were mentioned:

    • Ensure the database is correctly synchronized, so all charts and news reflect up-to-date data;
    • Hide or disable non-functional elements, rather than showing inactive buttons or placeholders — this avoids user frustration.

Iteration & Refinement #

Implemented features based on feedback #

Taking into account the user feedback, the following improvements were made:

  • LSTM model improved.
  • Interface updated: now silver and zinc data are shown; buttons for predictions on silver and zinc are disabled when not available.
  • The database is properly configured.
  • All buttons are now either clickable or disabled; no unused or confusing buttons remain.

Performance & Stability #

Frontend and backend metrics:

  • Website availability: users can easily access the website via the provided link.
  • Interface: users can select metals of interest, explore both historical and predicted graphs, and view parsed news.
  • Backend: users see dynamic graphs where historical prices update automatically, predictions update dynamically, and news updates are reflected timely.

ML metrics

The following metrics were used: mean squared error, mean absolute error, and the mean absolute percentage error. The first metric captures significant errors of the predicted values, while the second metric identifies the absolute error values. Finally, the last metric can be used to evaluate accuracy of a model.

Documentation #

1. Technical Research Documentation #

  • Location: ml/reports/
  • Purpose: Academic foundation, methodology, and validation
    • Literature Review.pdf - Research foundation
    • market_research.pdf - Market analysis
    • indicators_research.pdf - Technical indicators
    • feature selection.pdf - Feature engineering
    • Model_Evaluation_Report.pdf - Performance validation
    • Model_Refinement_Report.pdf - Performance improvement

2. Interactive Analysis Documentation #

  • Location: ml/notebooks/
  • Purpose: Reproducible analytical workflows
    • Data visualization and baseline models
    • Time series analysis
    • Statistical distribution analysis
    • Regression model development
    • Forecasting framework implementation
    • Hyperparameter tuning

3. Component Documentation #

  • File: ml/Readme.md
  • Purpose: Script explanations and usage instructions
    • Data collection workflows
    • Model training procedures
    • Dependencies and setup

4. Dependency Documentation #

  • Files: backend/requirements.txt, ml/requirements.txt
  • Purpose: Environment reproducibility and version control

ML Model Refinement #

Firstly, the LSTM model was analyzed on output and input features. The input features were enriched by adding periodic temporal data related to day of the week, day of the month, and season of the year. The output features were limited to close, high, and low prices as well as volume.

On top of that, as proposed in the relevant literature, the Savitzky-Golay filter was added in the dataset preparation part. This procedure improved the results of both baseline and LSTM models.

During the next week, we plan to explore a new model in place of LSTM to find a better fit for the datasets. In addition, extra features can be explored.

Weekly commitments #

Individual contribution of each participant #

From Friday to Monday:

From Monday to Wednesday:

Plan for Next Week #

  • Train models for zinc and silver.
  • Proceed working with gold model.
  • Refine the database.
  • Start updating UI according to the feedback.

Confirmation of the code’s operability #

We confirm that the code in the main branch:

  • [✓] In working condition.
  • [✓] Run via docker-compose (or another alternative described in the README.md).