7.5 Alternative Data Pipelines & Evaluation

Basic Prerequisites: Statistics & Inference (correlations, regression, the t-statistic) and Data Sources & Corporate Data (timestamps, vendors, point-in-time hygiene). No prior ML needed. (these are the folder-level prerequisites for pages 02–06; page 01 states its own, smaller, entry requirements)

1. Intuition & Practical Objective

Traditional market data - prices, volumes, filings - is universally available. Every fund reads the same tape, so any edge it contains is arbitraged to near zero within seconds. Alternative data is everything else: the primary information that has not yet reached the other sources (López de Prado, AFML §2.2.4). Before Exxon reported, someone watched the tankers; before a retailer guided up, someone counted the parking lot, monitored the card spend, or scraped the web prices. Alt-data’s promise is information that arrives weeks before it is reflected in price - and its cost is that the data is dirty, irregularly timestamped, expensive, and easy to backtest into a fantasy.

The alt-data lifecycle has five stages - hold these in your head:

1. Ingest - pull raw vendor delivery (files, API, stream).
2. Store - persist with two timestamps: when the event happened and when you learned it.
3. Clean - reconcile entities (CIK/ticker↔vendor id), de-duplicate, handle restatements.
4. Feature - aggregate into a point-in-time signal aligned to a tradeable universe.
5. Evaluate - measure whether it predicts a forward return out of sample and without leakage.

This folder is the hub. It (a) gives the fast alt-data category lookup and the two evaluation numbers (IC and decay half-life) below - job #1 of this folder - and (b) routes you to six sub-pages that climb from raw intuition to vendor economics and signal combination.

The one-sentence essence. “Alternative data is only a signal if it is (i) primary - not yet in prices - and (ii) honestly timestamped - a value is usable only at its knowledge date, never its event date - and everything downstream is the discipline of measuring a weak, decaying correlation without leaking the future.”

Audience arc: the beginner learns what alt-data is and why it is different from the tape; the intermediate learns the pipeline and how to test a dataset honestly (IC, decay); the expert learns vendor economics, uniqueness, and signal combination. The sub-pages are ordered for exactly that climb.

2. Mathematical Ground Truth & Quick Lookup

Notation. $x_t$ = the alt-data signal at time $t$; $y_{t+1}$ = the forward return it is meant to predict; $t_E$ = event time (when the real-world fact occurred); $t_K$ = knowledge time (when the record entered your database); $\lambda$ = decay rate; $B$ = breadth (number of independent bets per year); $N$ = cross-section size.

The point-in-time identity (the single rule that separates a real backtest from a fantasy):

Aligning a record to its event date - the way vendors index “the last day of the report period” (AFML §2.2.1) - injects up to one full reporting lag of look-ahead.

Alpha decay & half-life. A proprietary edge decays as the data is resold and arbitraged:

The evaluation statistics (the “does this dataset have signal?” measures):

The Fundamental Law of Active Management (Grinold & Kahn) - why every unique signal is worth hunting:

The alt-data category lookup (job #1). Categories per AFML §2.2.4 (after Kolanovic & Krishnamachari 2017):

Critical caveat - freshness, not just cleanliness. Vendors differ in cost and in when the value becomes knowable. A card panel aggregated weekly and delivered on a 3-day lag means the freshest value you may legally use at $t$ is 3–10 days old. Fix the lag convention once and hold it constant; the discipline that formalizes it is 05 · Failure Modes (Data Sources).

Verified quick-reference (reproduced on the sub-pages; see §3):

3. Computational Implementation - the minimal honest evaluation

The whole point in ~20 lines: build an alt-data signal, and show the same signal scores a perfect IC when read at its event date and a realistic one when read point-in-time. numpy only; stdlib works too.

4. Failure Modes & First-Principles Breakdowns (hub signposts)

The full analysis lives in 05 · Failure Modes & Practice. In one line each:

1. Look-ahead / point-in-time violation - reading a value at its event date instead of its knowledge date ($t_K$) injects the reporting lag into the backtest; the first principle is that a dataset is a function of when you looked at it (the §3 run shows the IC is doubled by this one mistake).
2. Backfill & survivorship bias - vendors overwrite history with corrected values and keep only today’s survivors; both flatter the backtest while the true universe looked worse leading in.
3. Overfitting the evaluation - testing one dataset is innocent; testing thousands and keeping the best is the multiple-testing crime, and an alt-data signal at $N=60$ names has $\text{IC}$ noise $\approx 1/\sqrt{60}$ - comparable to the signal itself.

5. References

• López de Prado, Marcos: Advances in Financial Machine Learning (Wiley, 2018)
• Kolanovic, Marko & Krishnamachari, Rajesh T.: Big Data and AI Strategies: Machine Learning and Alternative Data Approach to Investing, J.P. Morgan Global Quantitative & Derivatives Strategy (2017)
• Guida, Tony: Big Data and Machine Learning in Quantitative Investment (Wiley, 2019)
• AIMA / SS&C Technologies: Casting the Net: How Hedge Funds Are Using Alternative Data (2017)
• Grinold, Richard C. & Kahn, Ronald N.: Active Portfolio Management
• Luo, Yin et al.: “Seven Sins of Quantitative Investing” (Deutsche Bank, 2014)

6. Connected Graph Bridges

• Foundational base: Statistics & Inference (correlation, IC, $t$-stat) · Financial ML Pitfalls & Low-SNR (why a weak true edge dies in the noise)
• Data provenance: Data Sources & Corporate Data (timestamps, vendors, PIT hygiene) · Feature Engineering & Labeling (turning the cleaned panel into a modelable signal + target)
• Evaluation discipline: Backtesting Hygiene (deflated Sharpe, selection bias) · Purged CV & Backtest Hygiene
• Sibling topic: Financial NLP & Transcripts (alt-data one important special case: text)
• Sub-pages (in-folder): 01 From Zero · 02 Alt-Data Landscape · 03 The Pipeline · 04 Evaluating Signal · 05 Failure Modes · 06 Advanced Extensions

Beginner: start at 01 · Practitioner: start at 05