feat: reusable technical-analysis indicators and signal history models (#109) (#113)

Phase 1 of the technical-analysis framework:

- macros/technical_analysis: window primitives, price transforms, and a
  finite-window exponentially weighted mean (project-standard EMA/RMA
  for BigQuery, documented convergence)
- technical_price_universe: unified adjusted OHLCV spine across sp500
  stocks (split-adjusted), sector/fixed-income/currency ETFs, major
  indices, and global markets, with OHLC consistency clamping
- technical_indicator_daily: SMA/EMA, RSI(Wilder), MACD, Bollinger,
  ATR/NATR, ADX/+DI/-DI, stochastic, ROC, Williams %R, CCI, OBV, MFI,
  Donchian, relative volume, z-score, 52w distance; warmup-gated NULLs
- technical_signal_events: jinja signal registry (14 signals) with
  setup/triggered/active/completed/expired state machine
- technical_signal_instances: forward returns at 1/5/10/21/63/126 bars,
  SPY-relative returns, MFE/MAE, worked labels (no lookahead outside
  this evaluation model)
- technical_current_setups + agent_technical_signal_setups/history:
  operational current-state and reliability surfaces
- deprecates SPY-only technical_signals (VIX columns not yet migrated)

Validated with dbt build in dev: 7 models + 56 tests, all passing.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Author: C00ldudeNoonan

dbt_project/macros/technical_analysis/ta_price_transforms.sql

@@ -0,0 +1,40 @@
+{#
+    Price-transform expressions (issue #109). These are plain column
+    expressions — no window functions — so they can be used at any stage.
+#}
+
+{% macro ta_typical_price(high='high', low='low', close='close') %}
+    (({{ high }} + {{ low }} + {{ close }}) / 3)
+{% endmacro %}
+
+
+{% macro ta_median_price(high='high', low='low') %}
+    (({{ high }} + {{ low }}) / 2)
+{% endmacro %}
+
+
+{% macro ta_weighted_close(high='high', low='low', close='close') %}
+    (({{ high }} + {{ low }} + 2 * {{ close }}) / 4)
+{% endmacro %}
+
+
+{% macro ta_true_range(high='high', low='low', prev_close='prev_close') %}
+    {#- prev_close is expected to be a precomputed LAG(close) column.
+        Falls back to plain high-low on the first bar (prev_close NULL),
+        matching TA-Lib behavior. -#}
+    GREATEST(
+        {{ high }} - {{ low }},
+        COALESCE(ABS({{ high }} - {{ prev_close }}), {{ high }} - {{ low }}),
+        COALESCE(ABS({{ low }} - {{ prev_close }}), {{ high }} - {{ low }})
+    )
+{% endmacro %}
+
+
+{% macro ta_daily_return(close='close', prev_close='prev_close') %}
+    SAFE_DIVIDE({{ close }} - {{ prev_close }}, NULLIF({{ prev_close }}, 0))
+{% endmacro %}
+
+
+{% macro ta_log_return(close='close', prev_close='prev_close') %}
+    SAFE.LN(SAFE_DIVIDE({{ close }}, NULLIF({{ prev_close }}, 0)))
+{% endmacro %}

dbt_project/macros/technical_analysis/ta_primitives.sql

@@ -0,0 +1,136 @@
+{#
+    Technical-analysis window primitives (issue #109).
+
+    Conventions:
+    - Every macro accepts partition_by / order_by instead of hardcoding
+      symbol or date. Default partition is the symbol-level grain of
+      technical_price_universe: source_universe, symbol, exchange.
+    - Warmup handling is the caller's responsibility: gate outputs with
+      `CASE WHEN bars_available >= <window> THEN ... END` so indicators
+      are NULL until enough lookback rows exist.
+#}
+
+{% macro ta_default_partition() %}
+    {{- return('source_universe, symbol, exchange') -}}
+{% endmacro %}
+
+
+{% macro ta_rolling(agg, column, window, partition_by=none, order_by='date') %}
+    {#- Rolling aggregate over the trailing `window` rows (inclusive of current). -#}
+    {%- set partition_by = partition_by or ta_default_partition() -%}
+    {{ agg }}({{ column }}) OVER (
+        PARTITION BY {{ partition_by }}
+        ORDER BY {{ order_by }}
+        ROWS BETWEEN {{ window - 1 }} PRECEDING AND CURRENT ROW
+    )
+{% endmacro %}
+
+
+{% macro ta_rolling_prior(agg, column, window, partition_by=none, order_by='date') %}
+    {#- Rolling aggregate over `window` rows ENDING at the prior row.
+        Used for breakout levels (e.g. Donchian) so today's bar cannot
+        confirm its own breakout. -#}
+    {%- set partition_by = partition_by or ta_default_partition() -%}
+    {{ agg }}({{ column }}) OVER (
+        PARTITION BY {{ partition_by }}
+        ORDER BY {{ order_by }}
+        ROWS BETWEEN {{ window }} PRECEDING AND 1 PRECEDING
+    )
+{% endmacro %}
+
+
+{% macro ta_lag(column, n=1, partition_by=none, order_by='date') %}
+    {%- set partition_by = partition_by or ta_default_partition() -%}
+    LAG({{ column }}, {{ n }}) OVER (
+        PARTITION BY {{ partition_by }}
+        ORDER BY {{ order_by }}
+    )
+{% endmacro %}
+
+
+{% macro ta_lead(column, n=1, partition_by=none, order_by='date') %}
+    {%- set partition_by = partition_by or ta_default_partition() -%}
+    LEAD({{ column }}, {{ n }}) OVER (
+        PARTITION BY {{ partition_by }}
+        ORDER BY {{ order_by }}
+    )
+{% endmacro %}
+
+
+{% macro ta_zscore(column, window, partition_by=none, order_by='date') %}
+    {%- set partition_by = partition_by or ta_default_partition() -%}
+    SAFE_DIVIDE(
+        {{ column }} - {{ ta_rolling('AVG', column, window, partition_by, order_by) }},
+        NULLIF({{ ta_rolling('STDDEV', column, window, partition_by, order_by) }}, 0)
+    )
+{% endmacro %}
+
+
+{% macro ta_range_position(column, window, partition_by=none, order_by='date') %}
+    {#- Position of the current value within the trailing rolling range:
+        0.0 at the rolling minimum, 1.0 at the rolling maximum. -#}
+    {%- set partition_by = partition_by or ta_default_partition() -%}
+    SAFE_DIVIDE(
+        {{ column }} - {{ ta_rolling('MIN', column, window, partition_by, order_by) }},
+        NULLIF(
+            {{ ta_rolling('MAX', column, window, partition_by, order_by) }}
+            - {{ ta_rolling('MIN', column, window, partition_by, order_by) }},
+            0
+        )
+    )
+{% endmacro %}
+
+
+{% macro ta_window_array(column, lookback, partition_by=none, order_by='date') %}
+    {#- Trailing window of values as an ARRAY of single-field STRUCTs,
+        oldest first, for use with ta_ewm_from_array. Values are wrapped
+        in a STRUCT because BigQuery arrays cannot hold NULL elements and
+        analytic ARRAY_AGG does not support IGNORE NULLS; NULL values are
+        filtered inside the consuming subquery instead, which keeps each
+        element's recency offset intact. -#}
+    {%- set partition_by = partition_by or ta_default_partition() -%}
+    ARRAY_AGG(STRUCT({{ column }} AS v)) OVER (
+        PARTITION BY {{ partition_by }}
+        ORDER BY {{ order_by }}
+        ROWS BETWEEN {{ lookback - 1 }} PRECEDING AND CURRENT ROW
+    )
+{% endmacro %}
+
+
+{% macro ta_ewm_from_array(array_col, alpha) %}
+    {#- Exponentially weighted mean of a trailing-window array built by
+        ta_window_array (oldest element first, newest last).
+
+        This is the project-standard EMA implementation for BigQuery
+        (issue #109 open question): a finite-window, weight-normalized
+        approximation of the recursive EMA. With a lookback of >= 4/alpha
+        rows the truncated tail carries < 2% of total weight, so results
+        match the recursive definition to well within float noise.
+
+        Weights are proportional to (1-alpha)^(-offset); dividing by the
+        weight sum makes the constant factor cancel, so no array-length
+        term is needed.
+
+        EMA(span):  alpha = 2 / (span + 1)
+        RMA/Wilder(n): alpha = 1 / n
+    -#}
+    (
+        SELECT
+            SAFE_DIVIDE(
+                SUM(_e.v * POW({{ 1 - alpha }}, -_off)),
+                NULLIF(SUM(POW({{ 1 - alpha }}, -_off)), 0)
+            )
+        FROM UNNEST({{ array_col }}) AS _e WITH OFFSET AS _off
+        WHERE _e.v IS NOT NULL
+    )
+{% endmacro %}
+
+
+{% macro ta_ewm_mean_abs_dev_from_array(array_col, reference) %}
+    {#- Mean absolute deviation of array values from a fixed reference
+        (used by CCI, which measures dispersion around the CURRENT SMA). -#}
+    (
+        SELECT AVG(ABS(_e.v - {{ reference }}))
+        FROM UNNEST({{ array_col }}) AS _e
+    )
+{% endmacro %}

dbt_project/models/agents_preprocess/agent_technical_signal_history.sql

@@ -0,0 +1,56 @@
+{{ config(
+    description='Agent-friendly reliability summary of historical technical signals by universe, symbol, and signal (issue #109, phase 1)'
+) }}
+
+/*
+    Agent Technical Signal History
+
+    Aggregates technical_signal_instances into signal-reliability stats
+    at two grains (flagged by aggregation_grain):
+    - 'universe_signal':        source_universe × signal
+    - 'universe_symbol_signal': source_universe × symbol × signal
+
+    Hit rates are benchmark-relative (see worked_* definitions in
+    technical_signal_instances) and computed only over instances old
+    enough to be evaluable at each horizon.
+*/
+
+{% set grains = [
+    {'label': 'universe_signal', 'symbol_expr': "'ALL'", 'exchange_expr': "'ALL'"},
+    {'label': 'universe_symbol_signal', 'symbol_expr': 'symbol', 'exchange_expr': 'exchange'},
+] %}
+
+{% for grain in grains %}
+SELECT
+    '{{ grain.label }}' AS aggregation_grain,
+    source_universe,
+    {{ grain.symbol_expr }} AS symbol,
+    {{ grain.exchange_expr }} AS exchange,
+    indicator_name,
+    signal_name,
+    signal_side,
+    COUNT(*) AS total_triggers,
+    MIN(entry_date) AS first_trigger_date,
+    MAX(entry_date) AS last_trigger_date,
+    COUNTIF(worked_21d IS NOT NULL) AS evaluable_21d,
+    ROUND(AVG(CAST(worked_5d AS INT64)), 4) AS hit_rate_5d,
+    ROUND(AVG(CAST(worked_21d AS INT64)), 4) AS hit_rate_21d,
+    ROUND(AVG(CAST(worked_63d AS INT64)), 4) AS hit_rate_63d,
+    ROUND(AVG(forward_return_21d), 6) AS avg_forward_return_21d,
+    ROUND(AVG(relative_forward_return_21d), 6) AS avg_relative_return_21d,
+    ROUND(AVG(relative_forward_return_63d), 6) AS avg_relative_return_63d,
+    ROUND(AVG(max_favorable_excursion_21d), 6) AS avg_mfe_21d,
+    ROUND(AVG(max_adverse_excursion_21d), 6) AS avg_mae_21d,
+    ROUND(AVG(CASE WHEN volume_confirmed THEN CAST(worked_21d AS INT64) END), 4)
+        AS hit_rate_21d_volume_confirmed
+FROM {{ ref('technical_signal_instances') }}
+GROUP BY
+    source_universe,
+    {% if grain.label == 'universe_symbol_signal' %}symbol, exchange,{% endif %}
+    indicator_name,
+    signal_name,
+    signal_side
+{% if not loop.last %}
+UNION ALL
+{% endif %}
+{% endfor %}

dbt_project/models/agents_preprocess/agent_technical_signal_setups.sql

@@ -0,0 +1,35 @@
+{{ config(
+    description='Agent-friendly view of current technical signal setups and active signals (issue #109, phase 1)'
+) }}
+
+/*
+    Agent Technical Signal Setups
+
+    One row per symbol × signal currently in a setup/triggered/active
+    state, phrased for LLM-agent consumption: descriptive state text and
+    the key numbers an agent needs to reason about the setup.
+*/
+
+SELECT
+    source_universe,
+    symbol,
+    exchange,
+    date AS as_of_date,
+    signal_name,
+    indicator_name,
+    signal_side,
+    signal_state,
+    CASE signal_state
+        WHEN 'setup' THEN 'Pre-trigger condition present; signal has not fired yet'
+        WHEN 'triggered' THEN 'Signal fired on the most recent bar'
+        WHEN 'active' THEN 'Signal fired recently and is still inside its holding window'
+    END AS state_description,
+    signal_value,
+    close AS last_close,
+    relative_volume,
+    volume_confirmed,
+    trigger_date,
+    setup_date,
+    bars_since_trigger,
+    max_holding_bars
+FROM {{ ref('technical_current_setups') }}

dbt_project/models/agents_preprocess/technical_agents_schema.yml

@@ -0,0 +1,50 @@
+version: 2
+
+models:
+  - name: agent_technical_signal_setups
+    description: >
+      Agent-friendly view of symbols currently in a technical setup,
+      freshly triggered, or inside an active signal window (issue #109
+      phase 1). One row per symbol × signal at the symbol's latest bar,
+      with descriptive state text for LLM consumption.
+    columns:
+      - name: signal_state
+        tests:
+          - not_null
+          - accepted_values:
+              arguments:
+                values: ['setup', 'triggered', 'active']
+      - name: as_of_date
+        tests:
+          - not_null
+
+  - name: agent_technical_signal_history
+    description: >
+      Agent-friendly reliability summary of historical technical signals
+      (issue #109 phase 1): trigger counts, benchmark-relative hit rates
+      at 5/21/63 trading days, average relative returns, and excursion
+      stats, at universe×signal and universe×symbol×signal grains.
+    tests:
+      - unique_combination:
+          arguments:
+            combination_of_columns:
+              ['aggregation_grain', 'source_universe', 'symbol', 'exchange', 'signal_name']
+    columns:
+      - name: aggregation_grain
+        tests:
+          - not_null
+          - accepted_values:
+              arguments:
+                values: ['universe_signal', 'universe_symbol_signal']
+      - name: total_triggers
+        tests:
+          - not_null
+          - value_in_range:
+              arguments:
+                min_value: 1
+      - name: hit_rate_21d
+        tests:
+          - value_in_range:
+              arguments:
+                min_value: 0
+                max_value: 1

dbt_project/models/analysis/technical_schema.yml

@@ -0,0 +1,41 @@
+version: 2
+
+models:
+  - name: technical_signal_instances
+    description: >
+      Historical evaluation of every triggered technical signal (issue
+      #109 phase 1). One row per source_universe, symbol, exchange,
+      signal_name, entry_date. Forward returns at 1/5/10/21/63/126 trading
+      days, SPY benchmark and relative returns, max favorable/adverse
+      excursion over 21 bars, and benchmark-relative worked_* labels.
+      Forward-looking data exists ONLY here — never in events/setups.
+      Recent triggers keep NULL forward fields until enough bars elapse.
+    tests:
+      - unique_combination:
+          arguments:
+            combination_of_columns:
+              ['source_universe', 'symbol', 'exchange', 'signal_name', 'entry_date']
+    columns:
+      - name: entry_date
+        tests:
+          - not_null
+      - name: entry_price
+        tests:
+          - not_null
+          - value_in_range:
+              arguments:
+                min_value: 0
+      - name: signal_side
+        tests:
+          - not_null
+          - accepted_values:
+              arguments:
+                values: ['bullish', 'bearish']
+      - name: max_adverse_excursion_21d
+        description: >
+          Lowest forward low vs entry over the next 21 bars. Usually
+          negative, but can be positive if price never trades below entry.
+      - name: max_favorable_excursion_21d
+        description: >
+          Highest forward high vs entry over the next 21 bars. Usually
+          positive, but can be negative if price gaps down and stays there.