Update src/streamlit_app.py

src/streamlit_app.py

@@ -8,6 +8,7 @@ import pandas as pd
 import numpy as np
 import matplotlib.pyplot as plt
 from numpy.lib.stride_tricks import sliding_window_view
+from scipy.stats import spearmanr
 from sklearn.ensemble import GradientBoostingClassifier
 from huggingface_hub import HfApi, hf_hub_download
 
@@ -725,6 +726,91 @@ def render_sniper_tab():
                         render_sniper_dashboard(pick, result)
 
 
+# =========================================================
+# ANALOG PATH PROJECTION (used by the Signal Tracker chart)
+# =========================================================
+
+ANALOG_LOOKBACK = 60
+ANALOG_FORWARD = 60
+ANALOG_TOP_N = 5
+ANALOG_WEIGHT_STRENGTH = 5
+
+
+def _similarity_matches(close, end_idx, lookback, forward):
+    if end_idx - lookback < 0:
+        return []
+    recent = close.iloc[end_idx - lookback:end_idx]
+    recent_norm = (recent / recent.iloc[0]).values.flatten()
+
+    matches = []
+    max_search = end_idx - lookback - forward
+    for i in range(max_search):
+        cand = close.iloc[i:i + lookback]
+        cand_norm = (cand / cand.iloc[0]).values.flatten()
+        corr = spearmanr(recent_norm, cand_norm)[0]
+        if not np.isnan(corr):
+            matches.append((float(corr), i))
+    return matches
+
+
+@st.cache_data(ttl=24 * 3600, show_spinner="Building analog path projection...")
+def compute_analog_projection(ticker: str, flag_date_str: str, lookback=ANALOG_LOOKBACK,
+                               forward=ANALOG_FORWARD, top_n=ANALOG_TOP_N,
+                               weight_strength=ANALOG_WEIGHT_STRENGTH):
+    hist = get_sniper_ticker_data(ticker)
+    if hist.empty:
+        return None
+    close = hist["Close"]
+    flag_date = pd.Timestamp(flag_date_str)
+    end_idx = int(close.index.searchsorted(flag_date))
+    end_idx = min(end_idx, len(close) - 1)
+
+    if end_idx - lookback - forward < 5:
+        return None  # not enough pre-flag history for this method
+
+    matches = _similarity_matches(close, end_idx, lookback, forward)
+    if not matches:
+        return None
+
+    matches.sort(key=lambda x: x[0], reverse=True)
+    filtered, used = [], []
+    for corr, idx in matches:
+        if all(abs(idx - u) >= lookback for u in used):
+            filtered.append((corr, idx))
+            used.append(idx)
+        if len(filtered) >= top_n:
+            break
+    if not filtered:
+        return None
+
+    corrs = np.array([c for c, _ in filtered])
+    weights = np.exp(weight_strength * corrs)
+    weights /= weights.sum()
+
+    anchor_price = close.iloc[end_idx]
+    projections = []
+    for corr, idx in filtered:
+        fwd = close.iloc[idx + lookback: idx + lookback + forward]
+        fwd_norm = (fwd / fwd.iloc[0]).values.flatten()
+        projections.append(fwd_norm * anchor_price)
+    projections = np.vstack(projections)
+
+    weighted_mean = np.average(projections, axis=0, weights=weights)
+    weighted_var = np.average((projections - weighted_mean) ** 2, axis=0, weights=weights)
+    weighted_std = np.sqrt(weighted_var)
+
+    future_dates = pd.bdate_range(start=close.index[end_idx], periods=forward + 1)[1:]
+
+    return {
+        "future_dates": future_dates,
+        "weighted_mean": weighted_mean,
+        "upper": weighted_mean + weighted_std,
+        "lower": weighted_mean - weighted_std,
+        "robust_similarity": float(np.sqrt(np.sum(weights * corrs ** 2)) * 100),
+        "match_count": len(filtered),
+    }
+
+
 # =========================================================
 # TAB 4: SIGNAL TRACKER (predicted vs. actual, out-of-sample)
 # =========================================================
@@ -829,10 +915,21 @@ def render_signal_tracker():
                 ax.axvline(flag_date, color="red", ls="-", alpha=0.4, label="Flagged here")
                 if pd.notna(flag_price):
                     ax.axhline(flag_price, color="gray", ls=":", alpha=0.6, label="Price at flag")
-                    if pd.notna(row.get("Avg_3M")):
-                        predicted_price = flag_price * (1 + row["Avg_3M"] / 100)
-                        ax.axhline(predicted_price, color="green", ls="--", alpha=0.7,
-                                   label="Predicted avg 3M level")
+
+                projection = compute_analog_projection(ticker, str(flag_date.date()))
+                if projection:
+                    ax.plot(projection["future_dates"], projection["weighted_mean"],
+                             color="green", ls="--", linewidth=2, label="Analog-weighted projection")
+                    ax.fill_between(projection["future_dates"], projection["lower"], projection["upper"],
+                                     color="green", alpha=0.15, label="Projection band (1σ)")
+                    st.caption(
+                        f"Projection built from the {projection['match_count']} most similar non-overlapping "
+                        f"60-day windows in {ticker}'s own history, as known on the flag date "
+                        f"(robust similarity: {projection['robust_similarity']:.1f}%). Read this as 'what "
+                        f"comparable past setups did,' not a calibrated forecast — a handful of historical "
+                        f"echoes, not independent confirming evidence."
+                    )
+
                 ax.set_title(f"{ticker} — flagged {flag_date.date()}")
                 ax.legend(loc="best", fontsize=9)
                 ax.grid(True, alpha=0.1)
@@ -858,4 +955,4 @@ with tab3:
     render_sniper_tab()
 
 with tab4:
-    render_signal_tracker() 
+    render_signal_tracker()