"I am not the market's memory. I am what the market forgets."
.---.
/ \
| () () |
\ ^ /
|||||
[STYX SHADE]
Ghost is a Family 5 — Pattern Recognition / Similarity-Based agent, working alongside its siblings Mneme and Trace. While Mneme searches the memory of large-cap SPY and Trace follows the thread of blue-chip DIA, Ghost haunts the edges of the small-cap universe, trading the **Russell 2000 (IWM)**.
Small-cap stocks encode more stress per bar than their large-cap counterparts. They twitch on liquidity shocks, react violently to sector rotations, and print distinct structural patterns. However, these patterns are also more volatile and degrade faster across regimes. To capture this unique behavior, Ghost runs a similarity-based KNN search using a lower cosine similarity gate (≥ 0.40) and focuses on the K=15 nearest neighbors to identify when historical small-cap stress patterns are repeating.
Named for the shades of the River Styx, who linger at the boundaries of memory and time. Ghost does not chase trend momentum or wait for standard indicators to line up. It watches the quiet, volatile twitches of small caps in the morning and midday, looking for the faint outlines of past market sessions.
Because small-cap memory is shorter and more fragile, Ghost does not require the absolute clarity that Mneme demands. It is willing to act on fainter echoes (similarity ≥ 0.40) because a subtle pattern match in small caps carries as much structural information as a highly defined pattern in large caps.
Ghost's job: detect the faint, fleeting footprints in small-cap time series, verify that they rhyme with historical stress periods, and strike before the memory fades.
During its developmental sweeps, Ghost's framework discovered a crucial microstructural truth: **the midday transition in small-cap stocks is pure noise**. While its large-cap siblings trade three windows, Ghost dropped Window B (the morning-to-afternoon transition) entirely after backtests revealed it was toxic to performance. Ghost operates two highly focused independent windows:
| Window | Observation Period | Prediction Period | Entry Time | Hard Flat |
|---|---|---|---|---|
| Window A (Morning) | 9:30 – 11:30 (24 bars) | 12:00 – 14:00 (2 hours) | 11:30 ET | 14:00 ET |
| Window C (Closing) | 12:00 – 14:00 (24 bars) | 14:00 – 15:30 (1.5 hours) | 14:00 ET | 15:30 ET |
Each window is independent. If both windows signal a setup on the same day, Ghost takes only the single strongest signal (highest consensus vote) to protect the desk's capital.
1. Permissive Cosine Similarity Gate (≥ 0.40): The closest historical neighbor must have a similarity score of at least 0.40. This is calibrated for the higher noise and faster decay of small-cap patterns.
2. Decisive Majority Vote Gate (≥ 62%): At least 10 of the 15 historical twins must have closed positive in their prediction segments. While the similarity threshold is lower, the directional consensus must remain absolute.
1. Invalidation Stop: A hard stop is placed at 1.5× ATR below the entry price. There is no trailing stop — if the level is broken, the pattern is invalidated.
2. Time-Based Flat: At the end of the window's prediction segment, the position is closed at market. Discretionary holding is strictly prohibited.
3. Macro Blackout: Ghost skips all FOMC, CPI, NFP, and PCE days, as macro shocks overwrite the delicate, high-dimensional patterns it relies upon.
Ghost's developmental data revealed that its strongest edge lies in **Window C (Lunch-to-Close)**. While Window A suffered from morning noise, Window C achieved a robust **1.54 Profit Factor** in out-of-sample testing, proving that the midday-to-afternoon transition in small caps contains a highly stable, repeatable structural pattern.