factorlens 0.2.8

Factor attribution and analytics CLI

FactorLens

A Rust CLI for deterministic factor attribution and analytics workflows, with optional AI-generated explanations.

FactorLens explains why metrics change by decomposing results into factor contributions and producing structured analysis artifacts. FactorLens follows a math-first, AI-second approach: deterministic analytics produce the artifacts, and the LLM layer interprets them.

Quick Example

Given a dataset with sales metrics:

factorlens analyze data/sales_100.csv
factorlens analyze-compare data/sales_100.csv data/sales_150.csv
factorlens explain-analyze

Example output (truncated):

Executive Delta

Top-5 concentration changed from 12.0% → 21.3% (+9.3 pp)
Segment count changed from 62 → 60 (-2)

Top Concentration Changes
1. US | Direct | Premium | 0   9 → 7 records
2. US | Direct | Core | 1      7 → 7 records
3. US | Direct | Core | 0      7 → 7 records
...

Design Principles

FactorLens follows a few simple design rules:

• Math-first, AI-second – deterministic factor attribution produces the artifacts, AI only explains them.
• CLI-first workflows – designed to run locally, in scripts, or inside pipelines.
• Structured outputs – results can be exported as Markdown, JSON, or HTML for humans and automation.
• Composable commands – analysis, comparison, and explanation steps can be combined in workflows.

What It Looks Like

cargo run -p factor_cli -- analyze \
  --input data/your_file.csv \
  --group-by region,product_line,channel \
  --metrics revenue_usd \
  --out artifacts/analysis.md

Example report excerpt:

## Executive Summary

- Largest segment is `US | Core | Direct` with 28.4% of records and 32.1% of total revenue_usd.
- Top 5 segments represent 61.5% of records and 67.9% of revenue_usd.

AI layer on top:

Summary:
Growth is concentrated in US direct channel performance, while product-line mix
is creating downside concentration risk in a small number of segments.

Workflow

Command	Purpose
analyze	factor/segment attribution from CSV or Postgres
analyze-suggest	infer likely dimensions/metrics/date and generate starter profile TOML
analyze-compare	snapshot delta analysis (biggest movers)
explain-analyze	executive narrative and actions from computed JSON
factors fit / factors regress	statistical factors (PCA) or known-factor regression

2-Minute Quickstart

# 1) baseline snapshot (100 rows)
factorlens analyze \
  --input data/factorlens_demo_sales_100.csv \
  --group-by region,channel,product_line,plan_tier \
  --metrics revenue_usd,cost_usd,orders \
  --rank-by revenue_usd \
  --out artifacts/demo_sales_100.md

# 2) new snapshot (150 rows)
factorlens analyze \
  --input data/factorlens_demo_sales_150.csv \
  --group-by region,channel,product_line,plan_tier \
  --metrics revenue_usd,cost_usd,orders \
  --rank-by revenue_usd \
  --out artifacts/demo_sales_150.md

# 3) compare + explain
factorlens analyze-compare \
  --base artifacts/demo_sales_100.json \
  --new artifacts/demo_sales_150.json \
  --output-format html \
  --out artifacts/demo_compare.html

factorlens explain-analyze \
  --backend bedrock \
  --model anthropic.claude-3-haiku-20240307-v1:0 \
  --analysis-json artifacts/demo_sales_150.json \
  --question "What are the top concentration risks and what 3 actions should we take in the next 30 days?"

One-command runner:

./scripts/demo_sales.sh
# optional Bedrock:
RUN_BEDROCK=1 AWS_REGION=eu-central-1 ./scripts/demo_sales.sh

Demo Data

Public-safe demo files included:

• data/factorlens_demo_sales_100.csv
• data/factorlens_demo_sales_150.csv (use for compare)

Optional Postgres load:

psql "$DATABASE_URL" -c "
create schema if not exists demo;
drop table if exists demo.factorlens_demo_sales_100;
drop table if exists demo.factorlens_demo_sales_150;
create table demo.factorlens_demo_sales_100 (
  order_date date,
  region text,
  channel text,
  product_line text,
  plan_tier int,
  revenue_usd numeric(14,2),
  cost_usd numeric(14,2),
  orders int
);
create table demo.factorlens_demo_sales_150 (like demo.factorlens_demo_sales_100);
"
psql "$DATABASE_URL" -c "\copy demo.factorlens_demo_sales_100 from 'data/factorlens_demo_sales_100.csv' with (format csv, header true)"
psql "$DATABASE_URL" -c "\copy demo.factorlens_demo_sales_150 from 'data/factorlens_demo_sales_150.csv' with (format csv, header true)"

Generate a starter profile automatically from a new dataset:

factorlens analyze-suggest \
  --input data/factorlens_demo_sales_150.csv \
  --out artifacts/demo_suggest.md \
  --profile-name demo_exec \
  --auto-group-k 4 \
  --max-metrics 3

Large file tip:

factorlens analyze-suggest \
  --input data/factorlens_demo_sales_150.csv \
  --out artifacts/demo_suggest_random.md \
  --sample-rows 1000 \
  --sample-mode random \
  --sample-seed 42

This writes:

• artifacts/demo_suggest.md (human summary)
• artifacts/demo_suggest.json (machine-readable suggestion report)
• artifacts/demo_suggest.toml (ready profile config block)

Architecture

flowchart LR
    A["CSV/Postgres"] --> B["Factor/Segment Model (Rust)"]
    B --> C["Attribution Artifacts (JSON/CSV)"]
    C --> D["Explanation Layer (Local LLM or Bedrock)"]
    C --> E["Reports (Markdown/HTML/JSON)"]

Math engine first, explanation layer second.

Why This Exists

Many analytics workflows produce dashboards without a clear explanation of why metrics changed. FactorLens prioritizes attribution and residual math first, then translates those computed results into business language.

What This Is Not

• Not a trading bot
• Not a price prediction model
• Not a chat-first analytics toy

FactorLens computes attribution first, then uses LLMs only to explain computed artifacts.

Integrations

• Local LLMs via llama.cpp
• AWS Bedrock
• Claude Desktop / Claude Code via MCP
• CSV and Postgres data sources

MVP Features

• Price ingestion from CSV
• PCA factor model fitting
• Portfolio factor attribution
• Residual outlier detection
• Artifact outputs (json + csv)
• Markdown report generation
• Explain command using a local llama.cpp backend (llama-cli) with a Bedrock-ready backend contract

Workspace Layout

• crates/factor_core: Returns, PCA, attribution math
• crates/factor_io: CSV IO and artifact writing
• crates/factor_cli: CLI binary (factorlens)
• crates/llm_local: LLMClient trait + local/bedrock backends
• crates/report: Markdown report generation

Build Instructions

For advanced build/release details, see BUILD_INSTRUCTIONS.md.

Quick local build:

cargo build -p factor_cli
cargo build -p factor_cli --release

Input Formats

prices.csv

• date (YYYY-MM-DD)
• ticker
• close

portfolio.csv (optional)

• ticker
• weight

holdings.csv (optional alternative to portfolio.csv)

• ticker
• either market_value or both shares and price

factors.csv (for known-factor regression mode)

• date (YYYY-MM-DD)
• one or more numeric factor columns (for example: MKT, SMB, HML)

Quick Start

cargo run -p factor_cli -- factors fit \
  --prices data/prices.csv \
  --k 3 \
  --out artifacts/ \
  --portfolio data/portfolio.csv

# safer residual analysis: auto-pick k (< number of assets)
cargo run -p factor_cli -- factors fit \
  --prices data/prices.csv \
  --k-auto \
  --out artifacts/ \
  --portfolio data/portfolio.csv

# alternative: derive weights automatically from holdings
cargo run -p factor_cli -- factors fit \
  --prices data/prices.csv \
  --k 3 \
  --out artifacts/ \
  --holdings data/holdings.csv

cargo run -p factor_cli -- report \
  --artifacts artifacts/ \
  --format markdown \
  --out artifacts/report.md

# known-factor regression mode (MKT/SMB/HML-style)
cargo run -p factor_cli -- factors regress \
  --prices data/prices.csv \
  --factors data/factors.csv \
  --out artifacts/ \
  --portfolio data/portfolio.csv

cargo run -p factor_cli -- explain \
  --backend local \
  --model models/llama.gguf \
  --artifacts artifacts/ \
  --question "What drove the largest drawdown?"

Notes

• explain --backend local expects llama-cli on your PATH.
• explain --backend bedrock uses AWS Bedrock via AWS CLI (aws bedrock-runtime converse).
• This project is designed for explainability of computed analytics, not market prediction.

Explainability Notes

• factors fit excludes weekend dates by default.
• Pass --include-weekends if your dataset intentionally includes weekend trading.
• explain supports focused analysis with --focus-factors.

Examples:

cargo run -p factor_cli -- factors fit --prices data/prices.csv --k 3 --out artifacts/ --portfolio data/portfolio.csv
cargo run -p factor_cli -- factors fit --prices data/prices.csv --k 3 --out artifacts/ --portfolio data/portfolio.csv --include-weekends

cargo run -p factor_cli -- explain --backend local --model models/llama_instruct.gguf --artifacts artifacts/ --question "What drove the largest drawdown?" --focus-factors factor_1,factor_2

Custom Factor Names

By default, FactorLens auto-generates factor names from your dataset loadings (top positive and negative loading tickers per factor), so it works on any dataset.

You can still override labels with a CSV or TSV file via --factor-labels.

Example data/factor_labels.csv:

factor,label
factor_1_contrib,Broad Market Beta
factor_2_contrib,Growth vs Value Rotation
factor_3_contrib,Idiosyncratic Spread

Use in explain:

cargo run -p factor_cli -- explain --backend local --model models/llama_instruct.gguf --artifacts artifacts/ --question "What drove the largest drawdown?" --factor-labels data/factor_labels.csv

Notes:

• Factor keys may be factor_1, factor_1_contrib, or just 1.
• # comment lines are ignored.

Suggested Questions

• What was the worst modeled drawdown day, and what factors drove it?
• On the worst day, what percentage came from each factor?
• Which factor is my largest average downside contributor over the full sample?
• Which dates had the biggest positive factor-driven gains?
• Which 5 days had the largest residuals (moves not explained by factors)?
• Did my risk concentration increase in the last month?
• Is my portfolio dominated by one factor or diversified across factors?
• How stable are exposures across time windows?
• Which factor changed direction most often?
• Which factor contributed most to volatility, not just returns?
• If I remove factor_1, how much modeled downside is left?
• Compare drawdown drivers with and without weekends included.
• Using only factor_1,factor_2, what drove the drawdown?
• Which assets are most aligned with factor_1 loadings?
• Which assets increased my exposure to downside factors most?

Generic Table Analysis

Analyze any CSV table by grouping columns and numeric metrics you choose:

cargo run -p factor_cli -- analyze \
  --input data/your_file.csv \
  --group-by region,product_line,channel \
  --out artifacts/analysis.md

# profile-based quick starts
cargo run -p factor_cli -- analyze \
  --input data/your_file.csv \
  --profile exec \
  --out artifacts/analysis_exec.md

cargo run -p factor_cli -- analyze \
  --input data/your_file.csv \
  --profile segment \
  --out artifacts/analysis_segment.md

cargo run -p factor_cli -- analyze \
  --input data/your_file.csv \
  --profile supplier \
  --out artifacts/analysis_supplier.md

# custom profile config (recommended for private/domain fields)
cargo run -p factor_cli -- analyze \
  --input data/your_file.csv \
  --profile exec_custom \
  --profile-config profiles/profiles.example.toml \
  --out artifacts/analysis.md

# filtered + ranked view
cargo run -p factor_cli -- analyze \
  --input data/your_file.csv \
  --where region=US \
  --rank-by revenue_usd \
  --agg median \
  --percentiles p50,p90 \
  --alert-top5-share 60 \
  --alert-blank-share 10 \
  --top 10 \
  --min-records 20 \
  --out artifacts/analysis_filtered_ranked.md

# text normalization for name/title grouping + JSON-only output
cargo run -p factor_cli -- analyze \
  --input data/your_file.csv \
  --group-by title \
  --metrics revenue_usd \
  --normalize-text-groups \
  --word-freq \
  --output-format html \
  --out artifacts/analysis_title.html

Auto-detect useful grouping columns (if --group-by is omitted):

cargo run -p factor_cli -- analyze \
  --input data/your_file.csv \
  --out artifacts/analysis_auto.md

Analyze Compare

Create two analysis snapshots, then compare them:

# base snapshot
cargo run -p factor_cli -- analyze \
  --input data/your_file_a.csv \
  --group-by region,channel,product_line \
  --metrics revenue_usd,cost_usd,orders \
  --rank-by revenue_usd \
  --out artifacts/analysis_a.md

# new snapshot
cargo run -p factor_cli -- analyze \
  --input data/your_file_b.csv \
  --group-by region,channel,product_line \
  --metrics revenue_usd,cost_usd,orders \
  --rank-by revenue_usd \
  --out artifacts/analysis_b.md

# compare (markdown)
cargo run -p factor_cli -- analyze-compare \
  --base artifacts/analysis_a.json \
  --new artifacts/analysis_b.json \
  --out artifacts/compare.md

# compare (html)
cargo run -p factor_cli -- analyze-compare \
  --base artifacts/analysis_a.json \
  --new artifacts/analysis_b.json \
  --output-format html \
  --out artifacts/compare.html

# compare (json)
cargo run -p factor_cli -- analyze-compare \
  --base artifacts/analysis_a.json \
  --new artifacts/analysis_b.json \
  --output-format json \
  --out artifacts/compare.json

# compare (both markdown + json)
cargo run -p factor_cli -- analyze-compare \
  --base artifacts/analysis_a.json \
  --new artifacts/analysis_b.json \
  --output-format both \
  --out artifacts/compare.md

Notes:

• analyze outputs <out>.json by default (--output-format both).
• analyze-compare supports --output-format md|html|json|both.
• --top-movers controls how many largest movers are shown (default: 10).

Or analyze directly from Postgres:

# option 1: inline query
factorlens analyze \
  --postgres-url "$DATABASE_URL" \
  --query "SELECT region, channel, revenue_usd, cost_usd FROM analytics.sales" \
  --postgres-ssl-mode require \
  --postgres-ca-file /path/to/rds-ca-bundle.pem \
  --profile exec_custom \
  --profile-config profiles/profiles.example.toml \
  --out artifacts/analysis.md

# option 2: query file
factorlens analyze \
  --postgres-url "$DATABASE_URL" \
  --query-file sql/sales_analysis.sql \
  --profile exec_custom \
  --profile-config profiles/profiles.example.toml \
  --out artifacts/analysis.md

# option 3: AWS RDS/Aurora TLS with explicit CA bundle (recommended in pods)
mkdir -p /home/jovyan/certs
curl -fL "https://truststore.pki.rds.amazonaws.com/global/global-bundle.pem" \
  -o /home/jovyan/certs/rds-global-bundle.pem

factorlens analyze \
  --query "SELECT * FROM schema.table_a LIMIT 5000" \
  --postgres-ssl-mode require \
  --postgres-ca-file /home/jovyan/certs/rds-global-bundle.pem \
  --profile exec_custom \
  --profile-config profiles/profiles.example.toml \
  --out artifacts/analysis.md

Notes:

• Outputs both markdown and JSON (<out>.json).
• If --metrics is omitted, numeric metrics are auto-detected from the input file.
• --profile built-ins (exec, segment, supplier) are generic (no hardcoded domain columns).
• Use --profile-config <path.toml> for your own private, file-specific profile mappings.
• Input source is exclusive: use either --input <csv> or --postgres-url + (--query or --query-file).
• --postgres-url can be omitted if DATABASE_URL env var is set.
• --postgres-ssl-mode supports prefer (default), require, or disable.
• --postgres-ca-file optionally adds PEM CA certificates for DB TLS verification.
• For AWS RDS/Aurora in containers/pods, pass explicit RDS CA bundle via --postgres-ca-file if TLS handshake fails with system certs.
• Recommended layout: commit profiles/profiles.example.toml, keep private variants as profiles/*.local.toml or profiles/*.private.toml (gitignored).
• --where accepts comma-separated column=value filters (AND semantics).
• --rank-by ranks groups by a chosen metric (default ranking is by count).
• --agg controls metric aggregation: sum (default), mean, or median.
• --percentiles adds optional metric columns (p50, p90) per metric.
• --count-only disables numeric metric aggregation and reports concentration using records only.
• --exclude-blank-groups drops (blank) segment keys before ranking/reporting.
• --alert-top5-share and --alert-blank-share add threshold-based alerts to report output.
• --alert-rule adds custom rules (for example: top5_record_share_pct>60, blank_share_pct>10, segments<50). Quote rules containing < or > in shell commands, for example: --alert-rule 'segments<50,top5_record_share_pct>60'.
• --top controls how many groups are listed in the report.
• --top-insights adds deterministic Top Risks and Top Opportunities bullets to the report.
• --opportunity-min-records sets minimum records required for Top Opportunities candidates (default: 2).
• --normalize-text-groups normalizes group values for columns like name/title (lowercase + punctuation cleanup).
• --word-freq adds a Top Words section/counts for name/title-style grouping columns.
• --output-format supports md, json, both (default), or html.
• --min-records drops tiny segments before ranking (useful to avoid one-record outliers).
• analyze-suggest --out-profile <path.toml> writes a ready profile file directly.

Example --profile-config file:

[profiles.exec_custom]
group_by = ["region", "channel"]
metrics = ["revenue_usd"]
rank_by = "revenue_usd"
top = 12
min_records = 20
auto_group_k = 3

pip Package Usage

Install from PyPI:

For packaging/build/publish details, see BUILD_INSTRUCTIONS.md.

pip install --upgrade factorlens==0.1.3
factorlens --help

Local model:

factorlens explain \
  --backend local \
  --model /path/to/model.gguf \
  --artifacts /path/to/artifacts \
  --question "What drove the largest drawdown?"

Bedrock:

export AWS_REGION=us-east-1
factorlens explain \
  --backend bedrock \
  --model anthropic.claude-3-5-sonnet-20240620-v1:0 \
  --artifacts /path/to/artifacts \
  --question "What drove the largest drawdown?"

Explain from generic table analysis output (analysis.json):

Local model

factorlens explain-analyze \
  --backend local \
  --model /path/to/model.gguf \
  --analysis-json /path/to/analysis.json \
  --question "What are the top concentration risks and 3 actions?"

Bedrock

factorlens explain-analyze \
  --backend bedrock \
  --model anthropic.claude-3-haiku-20240307-v1:0 \
  --analysis-json /path/to/analysis.json \
  --question "What are the top concentration risks and 3 actions?"

MCP Server (Optional)

If you want to call FactorLens as tools from an MCP client, use:

• scripts/mcp/factorlens_mcp_server.py
• scripts/mcp/README.md

Quick start:

pip install mcp
python scripts/mcp/factorlens_mcp_server.py

What Bedrock Step Is Doing

factorlens explain --backend bedrock does not compute analytics. It only explains already-computed artifacts.

Step-by-step:

1. You run analytics first (factors fit or analyze) to produce artifacts.
2. explain loads artifact context (for factor mode: factors.json, attribution.csv, outliers.csv).
3. FactorLens builds a constrained prompt from that context.
4. FactorLens calls AWS Bedrock through AWS CLI (aws bedrock-runtime converse).
5. Bedrock returns plain-text explanation grounded in the provided artifact context.

Important:

• analyze command = pure Rust analytics, no LLM used.
• explain command = LLM narrative layer over artifacts.
• For table-analysis markdown (analysis.md), you can optionally call Bedrock directly with AWS CLI by passing report text as prompt.