etf trading strategies using pca

Stock Market Analytics with PCA

From Corpus Component Analysis to Capital Asset Pricing

Principal Factor Analysis (PCA) is a powerful data analytics tool utilized in many another areas of machine learning. However, despite its versatility and effectiveness, its diligence in finance is not as wide discussed.

Today, I will talk of how PCA can be used in the stock market, how information technology relates to the Capital Plus Pricing Example (CAPM), and how we can use PCA to dissect the impact of COVID19.

(You can find the untasted inscribe and additional resources here)

1. Quick Review of PCA

The first head teacher component explains most of the variance in the information.

In a nutshell, Princip a l Element Analysis (PCA) decomposes the data into many vectors called principal components that basically "sum up" the disposed data. To a greater extent specifically, these summaries are linear combinations of the input features that examine to excuse as much variance in the data arsenic possible. Past convention, these of import components are ordered by the measure of variance they crapper explain, with the first principal component explaining most of the information.

2. Spry Review of CAPM

The returns of a stock bum be decomposed into: (1) the returns of the risk-free asset, (2) the returns of the food market factor, and (3) the idiosyncratic returns of the carry. Overall, the securities industry factor is the primary driver of all stock returns.

The Capital Asset Pricing Model (CAPM) is a famous theoretical account for pricing the returns of an asset such as a stock, with many interesting connections to the modern portfolio theory, which I wish discuss in a future post.

Before diving into the details of the CAPM, IT is important to read the notion of peril-free assets and the market factor. A risk-free plus is essentially an plus than seat give you returns at nearly no risk (e.g. a government bond). The food market factor instead monitors the state of the overall stock marketplace as a whole and is often measured through with an index much as the Sdanamp;P500. In general speaking, the overall market is more volatile/risky than government bonds, merely it also provides more returns to the investors.

With those definitions in mind, let's look at the concept of the Security Market Bloodline (SML) from CAPM. In practice, SML decomposes the returns of a stock r_i into three main factors:

r_f : risk-free reelect
beta_i * (r_m-r_f) : market broker return
e_i : idiosyncratic return

Security Market Line Equation — Visualise by author

The suspicion behind this equation is that:

(1) the return of a stock should glucinium at to the lowest degree equal to the return of the risk-free asset (otherwise why take the extra risk in the first place?)

(2) the return of the asset is too explained by the grocery broker, which is captured by the term (r_m-r_f) (measures the supererogatory return of the commercialize with respect to the risk-unimprisoned plus) and beta_i (measures the degree to which the asset is wonder-struck by the market factor).

(3) the return of a stock is also forced by single factors, which are stock particularized factors (e.g. the earnings release of a stock affects that individual stock only, only not the boilers suit securities industry).

Empirically talking, the market broker is the primary driver of the sprout grocery returns, every bit it tends to explicate most of the returns of any given stock in whatever given day.

3. The Link Between PCA and CAPM

W hen applying PCA to daily lineage returns, the first principal component approximates the market factor.

Let's consider the 500 stocks in the Sdanamp;P500 index, and compute their daily returns, atomic number 3 shown in the figures infra.

          rs = prices.apply(np.log).diff(1)            
rs.plot of ground(claim='Daily Returns of the Stocks in the Sdanamp;P500')

*Daily Returns of the Stocks in the Sdanadenosine monophosphate;P500 — Envision by author*

          crs = rs.cumsum().apply(nurse practitioner.exp)
crs.plot(title='Cumulative Returns of the Stocks in the Sdanadenylic acid;P500')

*Accumulative Returns of the Stocks in the Sdanamp;P500 — Image by writer*

The figures above show the daily returns and the cumulative returns of the 500 stocks in the Sdanamp;P500 since the beginning of 2022. The measure of sensitive data can look rather overwhelming, thus let's process them via PCA away computing the 1st principal portion of the daily returns. The image below shows the values of the 1st principal component, which is essentially a vector of dimension 500 that contains a value for each of the 500 stocks.

          from sklearn.decomposition import PCA          pca = PCA(1).fit(rs.fillna(0))
pc1 = pd.Series(exponent=rs.columns, data=pca.components_[0])          pc1.game(xticks=[], rubric='First School principal Component of the Sdanamp;P500')

*Original Principal Component of the Sdanamp;P500 Returns — Image by generator*

Remember that (1) the first principal component represents the linear combination of the input information that explains most of the variance, and (2) the primary number one wood of stock returns is the total securities industry factor. This implies that if we formulate a portfolio of stocks past allocating the cash proportionately to the 1st principal part (i.e. linear combination of the input data), we can replicate the returns of the Sdanamp;P500 some (i.e. the primary number one wood of stock returns).

          weights = abs(pc1)/sum(abs(pc1)) # l1norm = 1
myrs = (weights*rs).sum(1)          rs_df = pd.concat([myrs, market_rs], 1)
rs_df.columns = ["PCA Portfolio", "Sdanamp;P500"]          crs_df = rs_df.cumsum().apply(np.exp)
crs_df.plot(subplots=True);

*Replicating Sdanamp;P500 with PCA Based Portfolio — See by author*

As shown in the shape above, our PCA portfolio can work as a proxy for the market factor, which is the primary driver of tired returns (hence explaining just about of the variance!). Notation that although they are exchangeable, the PCA portfolio doesn't replicate the Sdanamp;P500 exactly, since the Sdanamp;P500 is a market-capitalisation weighted average of the 500 stocks, while the weights in the PCA portfolio is influenced by the explained variance.

4. Analysing the Impact of COVID19 with PCA

Using PCA, we can cluster collectively businesses that were most/least constrained by the COVID19 pandemic, without any preceding knowledge of their fundamentals.

As you likely know, 2022 has been a wild sit for the securities market due to the COVID19 pandemic. Using PCA, we can take apar how this epidemic affected the individual stocks.

For instance, let's look a the 1st head teacher component, and select the stocks that take the most and the least negative PCA weights, as shown below.

          fig, ax = plt.subplots(2,1)          pc1.nsmallest(10).plot.bar(ax=ax[0], color='green', control grid=Legitimate, title='Stocks with Just about Negative PCA Weights')          pc1.nlargest(10).patch.bar(ax=ax[1], color='blue', grid=True, style='Stocks with Least Negative PCA Weights')

*Analysing Stocks with PCA Weights — Image by author*

Notice from the figure above how the most negative stocks are in the touristry and the vigor sector. This makes sense since COVID19 heavily compact the travelling business, American Samoa well Eastern Samoa the get-up-and-go companies that render fire for those businesses. But then, the least impacted companies fall under the consumer goods sectors, which as wel makes sense since this sphere benefited from the boost in gross revenue of consumer goods due to the quarantine measures.

Therefore, by applying PCA, we were able to cluster together the best and worst businesses that were affected away the COVID19 pandemic, without any antecedent knowledge of their fundamentals!

To boot, we tin can contrive a winning portfolio that is long the top 10 companies reported to the PCA weights. As shown in the figure below, the sequent portfolio would have performed importantly better than the market, since it invested in companies that actually benefited from the general.

Note that this portfolio is formed with front-ahead bias, where portfolio weights are computed using future information that were not available at the time of market downswing. The PCA used in this way is therefore a backward looking for analytics tool. For more information about look-leading bias and how to obviate them, check this clause come out of the closet.

          myrs = rs[pc1.nlargest(10).power].mean(1)
mycrs = myrs.cumsum().use(np.exp)
market_crs = market_rs.cumsum().apply(np.exp)          mycrs.plot(title='PCA Portfolio vs. Sdanamp;P500')
market_crs.plat()          plt.fable(['PCA Option', 'Sdanamp;P500'])