Lab Notes: Experiment record

English Content ExperimentPublished translation

StockPulse Self-Improvement Experiment — 2026-08-25 (Failure)

August 25, 2026About 2 min
stockpulseAI 실험자기개선예측분석

StockPulse AI Self-Improvement Experiment — Daily Analysis

Date: 2026-08-25


🧠 LLM Forecast Analysis

Forecast: Down / 6,600–6,700

Forecast summary: The previous day's combined foreign and institutional selling and large-cap weakness had pushed the KOSPI below 6,700, weakening the earlier bullish expectation. A sharp fall in top market-cap stocks such as Samsung Electronics added downside pressure, so the forecast expected the correction to continue at the open. Individual investors were still buying, but a reversal in foreign flows was identified as the key variable.

Actual KOSPI close: 6743
Actual direction: Up
Accuracy: 0.35/1.0

Analysis:
The forecast focused on the early selloff and incorrectly predicted the final closing direction (up). Institutional and individual bargain buying and a rebound in large caps produced an intraday recovery. The system underestimated how quickly market sentiment could recover after the previous day's selling.

Improvement:
The morning forecast should explicitly consider intraday volatility and rebound potential. It should analyze the possibility of a “weak-open, strong-close” pattern in which an early low is followed by a rebound, even after heavy selling on the previous day.


🤖 ML Model Forecast Analysis

Model: LGB + XGB + LSTM three-model ensemble (AUC=0.7755, Acc=0.7086)
Market bias: Down (-1.9%)
Bullish candidate: Entotek
Bearish candidates: Hana Financial Group, Shinhan Financial Group

Forecast: Down (-1.9%)

Evaluation: ❌ Failure
Accuracy: 0.0/1.0

Analysis:
The model expected the market to fall -1.9%, but the KOSPI rose +0.68%, so the direction was completely wrong. Despite heavy foreign net selling, institutional and individual buying and a rebound in selected sectors, including semiconductors, lifted the index. The model missed this weak-open, strong-close regime.

Improvement:
In addition to index direction, the model should include foreign and institutional flow data and weighted sector leaders as features. Volatility and momentum indicators should also be strengthened so the time-series model can learn V-shaped reversals after an intraday selloff.


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 after an early selloff, check institutional and individual buying and technical rebound signals in large caps. Evaluate the possibility of a weak-open, strong-close pattern or a V-shaped rebound. Make a combined judgment that considers intraday volatility and rebound capacity rather than only the direction.”

The change is applied from the next morning forecast.

🤖 ML: Feature engineering

New flow and sector features are added: foreign net buying/selling amount, institutional buying share, and the semiconductor sector index. Volatility indicators are also introduced.

The change is applied to the next data-collection pipeline.

This English page is part of DevSnack's English Content SEO/GEO Experiment. The Korean source remains the canonical editorial origin for this pilot.

Open the Korean source →