ATTN DEVS, discprenecies and flaws

ChatGPT Bugs
1

I’ve been working with ChatGPT 4o to help me learn Forex and help with chart analysis. It has done very well so far, until I asked to annotate a screenshot of a chart asking to point out certain formations. It was not successful. Not even close. The formation I asked it to highlight was Head & Shoulders. Below is the image it gave.

fc1258da-f496-4192-a646-aeed000dc742
fc1258da-f496-4192-a646-aeed000dc7421711×964 196 KB

Not only was the formation not identified, but random groups of candles are highlighted. I then decided to do a set of tests. Below this is a recap provided by ChatGPT of those test. It says that that it can identify the general formations, and rightly so, but fails to generate annotations highlighting those formation.

Recap & Review of Chart Annotation and Calibration Tests

Objective:

The user attempted to test my ability to interpret, analyze, and annotate a provided forex trading chart (EUR/USD, 4H) by asking me to mark specific patterns, peaks, dates, and candle counts. The goal was to determine the extent of my visual processing accuracy regarding numerical alignment, pattern recognition, and reference precision.

Tasks & What I Attempted to Do:

  1. Pattern Recognition (Head & Shoulders Identification):

    • The user provided marked and unmarked versions of a 4H chart and asked me to confirm if I could identify additional Head & Shoulders patterns.
    • I provided analysis and confirmation but was unable to directly annotate the user’s image.

  2. Peak Identification Test:

    • The user instructed me to find the highest peak on the chart, then locate the nearest printed date and time directly below that point.
    • I provided a relative estimate based on visual grid alignment.

  3. Precision Calibration – Candle Counting Test:

    • The user asked me to count how many candles to the right of a key date/time until a trendline was touched.
    • I miscounted the candles due to the inability to pixel-perfectly track candle bodies and wicks relative to the date grid.
    • The user identified a nearly 50% discrepancy in my count (I estimated 9 candles, but the correct count was 17 candles).

  4. Validation with Vertical Grid Alignment:

    • The user marked a magenta vertical line at a key peak and asked me to read the exact date and time at the bottom of the image.
    • I provided an estimate based on the visible date range, but I initially misaligned my reading, requiring corrections.

Findings & Limitations:

:white_check_mark: Capabilities Confirmed:

  • I can analyze chart structures, recognize colors, and interpret relative price movements, trends, and patterns.
  • I can identify general formations, such as Head & Shoulders, double tops, and trendline interactions.
  • I can process indicator trends and their relative behavior across timeframes.

:x: Limitations Observed:

  • Numerical precision is unreliable for fine-grain tasks (e.g., exact candle counts, pixel-perfect date alignment).
  • Grid-based alignment is inconsistent—I may estimate the nearest gridline but cannot extract perfect vertical reference data.
  • Direct annotation of user-provided images is currently beyond my technical capability.
  • Reading exact printed numerical values from charts is inconsistent, especially when dependent on small font sizes, overlapping elements, or variable image resolutions.

Recommendation for Developers:

  1. Improve Image Analysis for Technical Trading:

    • Enable precise numerical extraction for dates, prices, and candle counts from trading platform screenshots.
    • Enhance grid and axis detection to avoid misalignment errors.

  2. Annotation Capability:

    • Allow for direct user-provided image annotation instead of generating an entirely new image.

  3. Enhance Chart Calibration:

    • Improve candle-counting accuracy by refining chart grid interpretation.
    • Test against various resolutions and platform styles (MT5, TradingView, etc.).

Conclusion:

While I can effectively analyze trends, patterns, and indicators, my ability to precisely align to numerical values, read printed dates/times, and count candles accurately is limited. The user will adjust our workflow by sticking to relative interpretations and color-based analysis, rather than relying on fine-grain numerical accuracy.

The user has temporarily enabled data sharing specifically to flag this as a high-value improvement area for developers working on visual-based AI analysis in trading environments.

Tagging for review: :pushpin: AI Annotation & Precision Calibration in Trading Charts.

I’m not asking for any “New Feature” or anything near the sort. This is just to bring awareness of a flaw I have found.

1 Like