AI attention breakout analogs

This is the minimal script a human user would actually run: load data, call EchoTime, inspect the returned object.

from pathlib import Path

import pandas as pd
from echotime import compare_series, rolling_similarity

data_path = Path(__file__).resolve().parents[1] / "data" / "real_ai_attention_breakouts.csv"
df = pd.read_csv(data_path)

threads_report = compare_series(
    df["deepseek_cumulative"],
    df["threads_cumulative"],
    left_name="DeepSeek",
    right_name="Threads",
)
chatgpt_report = compare_series(
    df["deepseek_cumulative"],
    df["chatgpt_cumulative"],
    left_name="DeepSeek",
    right_name="ChatGPT",
)
windows = rolling_similarity(df["deepseek_cumulative"], df["threads_cumulative"], window=20, step=5)

print(threads_report.to_summary_card_markdown())
print(
    {
        "threads_similarity": round(threads_report.similarity_score, 3),
        "chatgpt_similarity": round(chatgpt_report.similarity_score, 3),
        "rolling_windows": len(windows),
    }
)

You should see one analog with consistently stronger Pearson, Spearman, and DTW evidence than the others, plus rolling windows that support the choice.

# EchoTime similarity summary

**Compared:** DeepSeek vs Threads

## Headline

DeepSeek vs Threads: Pearson r 0.94, Spearman rho 1.00, Kendall tau 1.00. The levels line up much more than the day-to-day changes, so the relationship is easier to defend as a broad shape analogy than as a local-dynamics match. The weakest agreement appears in derivative similarity, so timing or regime differences probably matter.

## Familiar statistics

| metric | value |
|---|---:|
| Pearson r | 0.936 |
| Spearman rho | 1.000 |
| Kendall tau | 1.000 |
| Best-lag Pearson r | 0.936 |
| Mutual info | 0.641 |
| First-difference r | 0.199 |

## Time-series-specific metrics

| plain-language label | score |
|---|---:|
| dtw similarity | 0.608 |
| trend similarity | 0.540 |
| spectral similarity | 0.464 |
| shape similarity | 0.431 |

## Recommended next actions

- Plot both series after z-score normalization to show the shared shape without scale differences.
- Run rolling or windowed similarity if you expect the relationship to change over time.
- Use structural-profile similarity when scales, frequencies, or observation modes differ too much for raw-shape comparison.
- For cumulative or monotonic inputs, compare first differences or daily increments before making an analog claim.

• DeepSeek vs Threads: Pearson 0.94 / Spearman 1.00 / Mutual info 0.64; against ChatGPT: Pearson 0.67 / Spearman 1.00 / Mutual info 0.47.
• The rolling view makes the analog choice inspectable instead of hiding disagreement behind one opaque number.
• This is a real analog-selection workflow on public data, not a hand-drawn breakout curve.