7.5.2 The Alt-Data Landscape

Basic Prerequisites: 01 · From Zero and Statistics & Inference.

---

1. Intuition & Practical Objective

This page is the map: what alternative datasets are, how they are categorized, and - the practical question every PM asks - is a given dataset worth anything to a portfolio? The objective is the framework that answers “worth anything”: the Fundamental Law of Active Management, which says the value of a signal is $\text{IR}=\text{IC}\sqrt{B}$, and its extension to combining signals, which says a new dataset’s worth is governed not by its own IC but by its uniqueness - how much of it is not already in the signals you trade.

The central reframing: alt-data rarely raises your IC; it raises your breadth. Almost no single dataset has a large IC (weak signals are the norm - see Financial ML Pitfalls). What a portfolio of unique signals does is multiply your number of independent bets $B$ - and $\sqrt{B}$ is where the leverage lives. Twenty genuinely-uncorrelated $IC=0.03$ signals beat one $IC=0.06$ signal, because $\text{IR}$ grows like $\sqrt{n}$ while IC grows like $n$ only in a fantasy where the signals are perfect.

---

2. Mathematical Ground Truth & Derivations

The Fundamental Law. For a signal with information coefficient IC and $B$ independent bets per period (rebalances × names, adjusted for correlations),

$$\boxed{\text{IR}=\text{IC}\cdot \sqrt{B}\cdot \text{TC}},$$

with transfer coefficient $\text{TC}\in [0,1]$ capturing implementation friction (constraints, costs, capacity). Breadth, not IC, is the lever most strategies actually pull - and alt-data is the main way to add breadth because each unique dataset is a new, weakly-correlated bet stream.

Combining $n$ signals with correlation. Let each of $n$ alphas have standalone IC and let $\rho$ be their common pairwise correlation. An equal-weight composite has IC and IR

$${\text{IC}}_{\text{comb}}=\text{IC}\cdot \sqrt{\frac{n}{1+(n-1)\rho }},{\text{IR}}_{\text{comb}}=\text{IC}\sqrt{\frac{n}{1+(n-1)\rho }}\sqrt{B}.$$

The two limits are the whole lesson: as $n\to \infty$ with $\rho =0$, ${\text{IC}}_{\text{comb}}\sim \text{IC}\sqrt{n}$ (unbounded - every new independent alpha helps); with $\rho >0$, the denominator gives the ceiling ${\text{IC}}_{\text{comb}}\to \text{IC}/\sqrt{\rho }$, a finite wall. Correlated alt-data saturates; unique alt-data compounds.

Required IC vs breadth. Setting $\text{IR}=1$ (a decent standalone strategy) and reading off IC:

$${\text{IC}}^{\star }=1/\sqrt{B}:B=12\Rightarrow 0.289,B=52\Rightarrow 0.139,B=252\Rightarrow \mathrm{0.063.}$$

A monthly strategy needs an enormous IC to “work” alone; a daily strategy needs a tiny one. This is why alt-data desks are obsessed with frequency and breadth.

Uniqueness (why the margin is in the residual). If a model already uses a factor $f$, a new signal $s$ contributes only its orthogonal part. Define uniqueness $U=1-R^2$ from regressing $s$ on $f$ (and the existing factor set):

$$U=1-R^2,R^2=\frac{\mathrm{Var}(\hat{s})}{\mathrm{Var}(s)},\hat{s}=\text{projection of }s\text{ on the existing factors}.$$

The marginal IR contribution of $s$ scales with $\sqrt{U}$, not with its raw IC. A signal with $\text{IC}=0.05$ and $U=0.18$ adds less than a signal with $\text{IC}=0.03$ and $U=0.99$.

---

3. Computational Implementation - breadth, correlation, and the ceiling

Stdlib only. It computes the Fundamental Law for a single alpha, the $\sqrt{n}$ compounding of independent alphas, the saturation wall when they are correlated, and the IC required for $\text{IR}=1$ at each breadth.

Read the two middle blocks together. With independent alphas, ten $IC=0.03$ signals combine to $\text{IC}=0.0949$ and $\text{IR}=1.506$ - the $\sqrt{n}$ compounding, real diversification. With correlated alphas ($\rho =0.7$), adding signals barely helps: $\text{IR}$ rises from $0.517$ ($n=2$) to only $0.564$ ($n=25$) - it hits the $\text{IC}/\sqrt{\rho }$ wall. The marginal value of an alt-data dataset is its correlation to what you already trade, not its standalone IC.

---

4. Failure Modes & First-Principles Breakdowns

1. “We have 500 datasets” - but they are all the same signal. Correlated alt-data (card panels from two vendors, three satellite providers of parking-lot counts) adds almost no breadth. The first principle: breadth counts independent bets, not tickers or feeds; ${\text{IC}}_{\text{comb}}$ saturates at $\text{IC}/\sqrt{\rho }$.
2. Backtesting the IC and ignoring capacity/transfer. A big IC in a tiny, illiquid name set with no capacity is worthless ($\text{TC}\ll 1$). The law has three factors; optimizing IC alone is optimizing $\frac{1}{3}$ of the problem.
3. The uniqueness illusion. A new signal with high raw IC but high correlation to your existing model (low $U$) is already priced into your P&L; adding it double-counts and inflates the backtest. Always compute the residual (06 · Advanced Extensions).
4. Frequency ≠ breadth. Rebalancing the same signal more often does not create independent bets when the signal is autocorrelated; $B$ must be an effective bet count after accounting for serial and cross-sectional correlation (Backtesting Hygiene).

---

5. References

• Grinold, Richard C. & Kahn, Ronald N.: Active Portfolio Management (2nd ed.)
• López de Prado, Advances in Financial Machine Learning, §2.2.4 (the alt-data taxonomy: individuals / business processes / sensors, after Kolanovic & Krishnamachari)
• Kolanovic & Krishnamachari, Big Data and AI Strategies (J.P. Morgan, 2017)
• Guida, Tony, Big Data and Machine Learning in Quantitative Investment (Wiley, 2019)
• AIMA / SS&C, Casting the Net (2017)

---

6. Connected Graph Bridges

• Back: 01 · From Zero · Index Hub
• Forward: 03 · The Pipeline → 04 · Evaluating Signal
• Framework source: Backtesting Hygiene (why breadth must be effective breadth, and the multiple-testing cost of hunting many signals) · Financial ML Pitfalls & Low-SNR
• Where the data comes from: Data Sources & Corporate Data