StockPulse Self-Improvement Experiment — 2026-08-21 (Success)
StockPulse AI Self-Improvement Experiment — Daily Analysis
Date: 2026-08-21
🧠 LLM Forecast Analysis
Forecast: Up / 6,900–7,000
Forecast summary: Net spot buying by foreign investors and institutions on the previous day, together with strength in large caps such as SK hynix and Samsung Electronics, drove a sharp KOSPI rebound. With expectations of easing U.S. long-term yields and a buying-sidecar activation during the session confirming upward momentum, the forecast expected the uptrend to continue and aimed for a break above 6,900.
Actual KOSPI close: 6913
Actual direction: Up
Accuracy: 0.85/1.0
Analysis:
The forecast direction (up) and target zone (a break above 6,900) matched the actual result (+0.88%, closing at 6912.95). However, although the KOSPI rose, the number of declining stocks exceeded the number of advancing stocks, creating a quantitative divergence. Continued foreign selling weakened the quality and durability of the rebound.
Improvement:
The system should evaluate not only the index direction but also the ratio of advancing stocks to declining stocks and the degree of large-cap dependence. It should also set a quantitative limit for rebounds while foreign selling continues and strengthen risk-management instructions that temper excessive optimism.
🤖 ML Model Forecast Analysis
Model: LGB + XGB + LSTM three-model ensemble (AUC=0.7691, Acc=0.7002)
Market bias: Down (-2.3%)
Bullish candidates: Entotek, Alteogen, KB Financial, SK Square, S
Forecast: Down (-2.3%)
Evaluation: ❌ Failure
Accuracy: 0.0/1.0
Analysis:
The model forecast a -2.3% decline, but the KOSPI rose +0.88%, so the direction was completely wrong. The model failed to represent the structural difference between a broad risk-off market and an index rising because large caps defended the benchmark.
Improvement:
The system should move from predicting only index returns to multi-target learning that considers stock-level flows and sector dispersion. It should also add the weighted index contribution of large caps as a feature to resolve the divergence between the index and individual stocks.
This analysis was produced by AI (Qwen3.5-35B) as an evaluation of forecast performance. It is not investment advice; it is an AI performance measurement experiment.
📝 Applied Improvement Actions
🧠 LLM: Prompt improvement
The morning analysis prompt now includes: “Even when the KOSPI rises, if declining stocks outnumber advancing stocks by 2:1 or more, treat the move as a possible large-cap-driven index distortion. Be cautious when raising the target, and include a warning that continued foreign net selling may weaken the rebound momentum.”
The change is applied from the next morning forecast.
🤖 ML: Feature engineering
New sector-level flow indicators and large-cap weighted contribution features are added alongside index-return prediction. A hybrid labeling strategy separates index and individual-stock predictions.
The change is applied to the next data-collection pipeline.