Markets are, in my view, mostly random. However, they're not completely random. Many small inefficiencies and patterns exist in markets which can be identified and used to gain slight edge on the market.
These edges are rarely large enough to trade in isolation - transaction costs and overhead can easily exceed the expected profits offered. But when we are able to combine many such small edges together, the rewards can be great.
In this article, I'll present a framework for blending together outputs from multiple models using a type of ensemble modeling known as stacked generalization. This approach excels at creating models which "generalize" well to unknown future data, making them an excellent choice for the financial domain, where overfitting to past data is a major challenge.
This post is the sixth and final installment in my tutorial series on applying machine learning to financial time series data. If you haven't already read the prior articles, you may want to do that before starting this one.
Ensemble Learning¶
Ensemble learning is a powerful - and widely used - technique for improving model performance (especially it's generalization) by combining predictions made by multiple different machine learning models. The idea behind ensemble learning is not dissimilar from the concept "wisdom of the crowd", which posits that the aggregated/consensus answer of several diverse, well-informed individuals is typically better than any one individual within the group.
In the world of machine learning, this concept of combining multiple models takes many forms. The first form appears within a number of commonly used algorithms such as Random Forests, Bagging, and Boosting (though this one works somewhat differently). Each of these algorithms takes a single base model (e.g., a decision tree) and trains many versions of that single algorithm on differing sets of features or samples. The resulting collection of trained models are often more robust out of sample because they're likely to be less overfitted to certain features or samples in the training data.
A second form of ensembling methods involves aggregating across multiple different model types (e.g., an SVM, a logistic regression, and a decision tree) - or with different hyperparameters. Since each learning algorithm or set of hyperparameters tends to have different biases, it will tend to make different prediction errors - and extract different signals - from the same set of data. Assuming all models are reasonably good - and that the errors are reasonably uncorrelated to one another - they will partially cancel each other out and the aggregated predictions will be more useful than any single model's predictions.
One particularly flexible approach to this latter type of ensemble modeling is "stacked generalization", or "stacking". In this post, I will walk through a simple example of stacked generalization applied to time series data. If you'd like to replicate and experiment with the below code, you can download the source notebook for this post by right-clicking on the below button and choosing "save link as"
Overview of Stacked Generalization¶
The "stacked generalization" framework was initially proposed by Wolpert in a 1992 academic paper. Since it was first proposed, stacked generalization (aka "stacking") has received a modest but consistent amount of attention from the ML research community.
Stacked generalization is an ensemble modeling technique. The core concept of stacked generalization is to generate a single, optimally robust prediction for a regression or classification task by (a) building multiple different models (with varying learning algorithms, varying hyperparameters, and/or different features) to make predictions then (b) training a "meta-model" or "blending model" to determine how to combine the predictions of each of these multiple models.
A nice way to visualize this (borrowed from documentation for Sebastian Rashka's excellent mlxtend package) is shown below. Each model R1 thru Rm is trained on historical data and used to make predictions P1 thru Pm. Those predictions then become the features used to train a meta-model to determine how to combine these predictions.
I think of this using an analogy. Imagine that there is a team of investment analysts whose manager has asked each of them to make earnings forecasts for the same set of companies across many quarters. The manager "learns" which analysts have historically been most accurate, somewhat accurate, and inaccurate. When future predictions are needed, the manager can assign greater and lesser (and in some cases, zero) weighting to each analyst's prediction.
It's clear why it's referred to as "stacked". But why "generalization"? The principal motivation for applying this technique is to achieve greater "generalization" of models to out-of-sample (i.e., unseen) data by de-emphasizing models which appear to be overfitted to the data. This is achieved by allowing the meta-model to learn which of the base models' predictions have held up well (and poorly) out-of-sample and to weight models appropriately.
Motivations¶
In my view, stacked generalization is perfectly suited to the challenges we face when making predictions in noisy, non-stationary, regime-switching financial markets. When properly implemented (see next section), stacking help to defend against the scourge of overfitting - something which virtually all practitioners of investing ML will agree is a major challenge.
Better yet, stacking allows us to blend together relatively weak (but orthogonal and additive) signals together in a way that doesn't get drowned out by stronger signals.
To illustrate, consider a canonical trend-following strategy which is predicated on 12 month minus 1 month price change. Perhaps we also believe that month-of-year or recent IBIS earnings trend have a weak, but still useful effect on price changes. If we were to train a model that lumped together dominant features (12 minus 1 momentum) and weaker features (seasonality or IBIS trend), our model may miss the subtle information because the dominant features overshadow them.
A stacked model, which has one component (i.e., a base model) focused on solely momentum features, another component focused on solely seasonality features, and a third one focused on analyst revisions features can capture and use the more subtle effects alongside the more dominant momentum effect.
Keys to Success¶
Stacked generalization is sometimes referred to as a "black art" and there is truth to that view. However, there are also two concrete principles that will get you a long way towards robust results.
1. Out of Sample Training
First, it's absolutely critical that the predictions P1 thru Pm used to train the meta-model are exclusively out of sample predictions. Why? Because in order to determine which models are likely to generalize best to out of sample (ie those with least overfit), we must judge that based on past predictions which were themselves made out-of-sample.
Imagine that you trained two models using different algorithms, say logistic regression and decision trees. Both could be very useful (out of sample) but decision trees have a greater tendency to overfit training data. If we used in-sample predictions as features to our meta-learner, we'd likely give much more weight to the model with a tendancy to overfit the most.
Several methods can be used for this purpose. Some advise splitting training data into Train1 and Train2 sets so base models can be trained on Train1 and then can make predictions on Train2 data for use in training the ensemble model. Predictions of the ensemble model must, of course, be evaluated on yet another dataset.
Others use K-fold cross-validation prediction (such as scikit's cross_val_predict) on base models to simulate out-of-sample(ish) predictions to feed into the ensemble layer.
However, in my view, the best method for financial time series data is to use walk-forward training and prediction on the base models, as described in my Walk-forward modeling post. In addition to ensuring that every base prediction is true out-of-sample, it simulates the impact of non-stationarity (a.k.a. regime change) over time.
2. Non-Negativity
Second - and this is less of a hard-and-fast rule - is to constrain the meta-model to learning non-negative coefficients only, using an algorithm like ElasticNet or lasso which allows non-negativity constraints.
This technique is important because quite often (and sometimes by design) there will be very high collinearity of the "features" fed into the meta-model (P1 thru Pm). In periods of high collinearity, learning algorithms can do funky things, such as finding a slightly better fit to past data by assigning a high positive coefficient to one model and a large negative coefficient to another. This is rarely what we really want.
Call me crazy, but if a model is useful only in that it consistently predicts the wrong outcome, it's probably not a model I want to trust.
Further Reading¶
That's enough (too much?) background for now. Those interested in more about the theory and practice of stacked generalization should check out the below research papers:
Preparing the Data¶
For this simple example, I will create synthetic data rather than using real market prices to remove the ambiguity about what features and transformations may be necessary to extract maximum value from the model.
Note: to make the dataset more realistic, I will extract an index from actual stock prices using quandl's API, but all features and target values will be constructed below.
With index in hand, we will generate four "hidden factors". These are the non-random drivers of the target variable, and are the "signal" we ideally want to learn.
To ensure that these factors are meaningful, we will create the target variable (y) using combinations of these factors. The first two hidden factors have a linear relationship to the target. The second two hidden factors have a more complex relationship involving interaction effects between variables. Lastly, we will add a noise component to make our learners work for it.
Finally, we'll create several features that are each related to one or more hidden factors, including generous amounts of noise and bias.
Key point: we've created X and y data which we know is related by several different linkages, some of which are linear and some of which aren't. This is what our modeling will seek to learn.
import numpy as np
import pandas as pd
pd.core.common.is_list_like = pd.api.types.is_list_like # remove once updated pandas-datareader issue is fixed
# https://github.com/pydata/pandas-datareader/issues/534
import pandas_datareader.data as web
%matplotlib inline
from IPython.core.display import HTML,Image
HTML('<style>{}</style>'.format(config.CSS))
def get_symbols(symbols,data_source, begin_date=None,end_date=None):
out = pd.DataFrame()
for symbol in symbols:
df = web.DataReader(symbol, data_source,begin_date, end_date)[['AdjOpen','AdjHigh','AdjLow','AdjClose','AdjVolume']].reset_index()
df.columns = ['date','open','high','low','close','volume'] #my convention: always lowercase
df['symbol'] = symbol # add a new column which contains the symbol so we can keep multiple symbols in the same dataframe
df = df.set_index(['date','symbol'])
out = pd.concat([out,df],axis=0) #stacks on top of previously collected data
return out.sort_index()
idx = get_symbols(['AAPL','CSCO','MSFT','INTC'],data_source='quandl',begin_date='2012-01-01',end_date=None).index
# note, we're only using quandl prices to generate a realistic multi-index of dates and symbols
num_obs = len(idx)
split = int(num_obs*.80)
## First, create factors hidden within feature set
hidden_factor_1 = pd.Series(np.random.randn(num_obs),index=idx)
hidden_factor_2 = pd.Series(np.random.randn(num_obs),index=idx)
hidden_factor_3 = pd.Series(np.random.randn(num_obs),index=idx)
hidden_factor_4 = pd.Series(np.random.randn(num_obs),index=idx)
## Next, generate outcome variable y that is related to these hidden factors
y = (0.5*hidden_factor_1 + 0.5*hidden_factor_2 + # factors linearly related to outcome
hidden_factor_3 * np.sign(hidden_factor_4) + hidden_factor_4*np.sign(hidden_factor_3)+ # factors with non-linear relationships
pd.Series(np.random.randn(num_obs),index=idx)).rename('y') # noise
## Generate features which contain a mix of one or more hidden factors plus noise and bias
f1 = 0.25*hidden_factor_1 + pd.Series(np.random.randn(num_obs),index=idx) + 0.5
f2 = 0.5*hidden_factor_1 + pd.Series(np.random.randn(num_obs),index=idx) - 0.5
f3 = 0.25*hidden_factor_2 + pd.Series(np.random.randn(num_obs),index=idx) + 2.0
f4 = 0.5*hidden_factor_2 + pd.Series(np.random.randn(num_obs),index=idx) - 2.0
f5 = 0.25*hidden_factor_1 + 0.25*hidden_factor_2 + pd.Series(np.random.randn(num_obs),index=idx)
f6 = 0.25*hidden_factor_3 + pd.Series(np.random.randn(num_obs),index=idx) + 0.5
f7 = 0.5*hidden_factor_3 + pd.Series(np.random.randn(num_obs),index=idx) - 0.5
f8 = 0.25*hidden_factor_4 + pd.Series(np.random.randn(num_obs),index=idx) + 2.0
f9 = 0.5*hidden_factor_4 + pd.Series(np.random.randn(num_obs),index=idx) - 2.0
f10 = hidden_factor_3 + hidden_factor_4 + pd.Series(np.random.randn(num_obs),index=idx)
## From these features, create an X dataframe
X = pd.concat([f1.rename('f1'),f2.rename('f2'),f3.rename('f3'),f4.rename('f4'),f5.rename('f5'),
f6.rename('f6'),f7.rename('f7'),f8.rename('f8'),f9.rename('f9'),f10.rename('f10')],axis=1)
Exploratory Data Analysis¶
We'll avoid going too deeply into exploratory data analysis (refer to Feature engineering for more exploration of that topic). I will, however, plot three simple views:
- Distributions of the features and target variables.
- Simple univariate regressions for each of the ten features vs. the target variable.
- A clustermap showing correlations between the features (see Feature selection for more on this technique):
## Distribution of features and target
X.plot.kde(legend=True,xlim=(-5,5),color=['green']*5+['orange']*5,title='Distributions - Features and Target')
y.plot.kde(legend=True,linestyle='--',color='red') # target
## Univariate Regressions
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
sns.set(style="dark")
# Set up the matplotlib figure
fig, axes = plt.subplots(4, 3, figsize=(8, 6), sharex=True, sharey=True)
# Rotate the starting point around the cubehelix hue circle
for ax, s in zip(axes.flat, range(10)):
cmap = sns.cubehelix_palette(start=s, light=1, as_cmap=True)
x = X.iloc[:,s]
sns.regplot(x, y,fit_reg = True, marker=',', scatter_kws={'s':1},ax=ax,color='salmon')
ax.set(xlim=(-5, 5), ylim=(-5, 5))
ax.text(x=0,y=0,s=x.name.upper(),color='black',
**{'ha': 'center', 'va': 'center', 'family': 'sans-serif'},fontsize=20)
fig.tight_layout()
fig.suptitle("Univariate Regressions for Features", y=1.05,fontsize=20)
## Feature correlations
from scipy.cluster import hierarchy
from scipy.spatial import distance
corr_matrix = X.corr()
correlations_array = np.asarray(corr_matrix)
linkage = hierarchy.linkage(distance.pdist(correlations_array), \
method='average')
g = sns.clustermap(corr_matrix,row_linkage=linkage,col_linkage=linkage,\
row_cluster=True,col_cluster=True,figsize=(5,5),cmap='Greens',center=0.5)
plt.setp(g.ax_heatmap.yaxis.get_majorticklabels(), rotation=0)
plt.show()
label_order = corr_matrix.iloc[:,g.dendrogram_row.reordered_ind].columns
Making Base Models¶
The first step in a stacked generalization system is to generate the "base models", meaning the models which are learning from our input features. We'll create two base models to use in our ensemble:
- A collection of simple linear regression models
- A collection of tree models - in this case, using the ExtraTrees algorithm
As described above it's absolutely critical to build models which provide realistic out-of-sample predictions, I am going to apply the methodology presented in Walk-forward modeling. In short, this will retrain at the end of each calendar quarter, using only data which would have been available at that time. Predictions are made using the most recently trained model.
To make this easier to follow, I'll define a simple function called make_walkforward_model that trains a series of models at various points in time and generates out of sample predictions using those trained models.
from sklearn.base import clone
from sklearn.linear_model import LinearRegression
def make_walkforward_model(features,outcome,algo=LinearRegression()):
recalc_dates = features.resample('Q',level='date').mean().index.values[:-1]
## Train models
models = pd.Series(index=recalc_dates)
for date in recalc_dates:
X_train = features.xs(slice(None,date),level='date',drop_level=False)
y_train = outcome.xs(slice(None,date),level='date',drop_level=False)
#print(f'Train with data prior to: {date} ({y_train.count()} obs)')
model = clone(algo)
model.fit(X_train,y_train)
models.loc[date] = model
begin_dates = models.index
end_dates = models.index[1:].append(pd.to_datetime(['2099-12-31']))
## Generate OUT OF SAMPLE walk-forward predictions
predictions = pd.Series(index=features.index)
for i,model in enumerate(models): #loop thru each models object in collection
#print(f'Using model trained on {begin_dates[i]}, Predict from: {begin_dates[i]} to: {end_dates[i]}')
X = features.xs(slice(begin_dates[i],end_dates[i]),level='date',drop_level=False)
p = pd.Series(model.predict(X),index=X.index)
predictions.loc[X.index] = p
return models,predictions
To create a series of walk-forward models, simply pass in X and y data along with a scikit estimator object. It returns a series of models and a series of predictions. Here, we'll create two base models on all features, one using linear regression and one with extra trees.
from sklearn.linear_model import LinearRegression
from sklearn.ensemble import ExtraTreesRegressor
linear_models,linear_preds = make_walkforward_model(X,y,algo=LinearRegression())
tree_models,tree_preds = make_walkforward_model(X,y,algo=ExtraTreesRegressor())
Note that no predictions can be made prior to the first trained model, so it's important to dropna() on predictions prior to use.
print("Models:")
print(linear_models.head())
print()
print("Predictions:")
print(linear_preds.dropna().head())
It can be instructive to see how the linear model coefficients evolve over time:
pd.DataFrame([model.coef_ for model in linear_models],
columns=X.columns,index=linear_models.index).plot(title='Weighting Coefficients for \nLinear Model')
Next, I'll create a simple function to evaluate multiple model performance metrics, called calc_scorecard. Further detail on this method of model evaluation is provided in the earlier post Model evaluation.
from sklearn.metrics import r2_score,mean_absolute_error
def calc_scorecard(y_pred,y_true):
def make_df(y_pred,y_true):
y_pred.name = 'y_pred'
y_true.name = 'y_true'
df = pd.concat([y_pred,y_true],axis=1).dropna()
df['sign_pred'] = df.y_pred.apply(np.sign)
df['sign_true'] = df.y_true.apply(np.sign)
df['is_correct'] = 0
df.loc[df.sign_pred * df.sign_true > 0 ,'is_correct'] = 1 # only registers 1 when prediction was made AND it was correct
df['is_incorrect'] = 0
df.loc[df.sign_pred * df.sign_true < 0,'is_incorrect'] = 1 # only registers 1 when prediction was made AND it was wrong
df['is_predicted'] = df.is_correct + df.is_incorrect
df['result'] = df.sign_pred * df.y_true
return df
df = make_df(y_pred,y_true)
scorecard = pd.Series()
# building block metrics
scorecard.loc['RSQ'] = r2_score(df.y_true,df.y_pred)
scorecard.loc['MAE'] = mean_absolute_error(df.y_true,df.y_pred)
scorecard.loc['directional_accuracy'] = df.is_correct.sum()*1. / (df.is_predicted.sum()*1.)*100
scorecard.loc['edge'] = df.result.mean()
scorecard.loc['noise'] = df.y_pred.diff().abs().mean()
# derived metrics
scorecard.loc['edge_to_noise'] = scorecard.loc['edge'] / scorecard.loc['noise']
scorecard.loc['edge_to_mae'] = scorecard.loc['edge'] / scorecard.loc['MAE']
return scorecard
calc_scorecard(y_pred=linear_preds,y_true=y).rename('Linear')
Since we are concerned about not only average performance but also period-to-period consistency, I'll create a simple wrapper function which recalculates our metrics by quarter.
def scores_over_time(y_pred,y_true):
df = pd.concat([y_pred,y_true],axis=1).dropna().reset_index().set_index('date')
scores = df.resample('A').apply(lambda df: calc_scorecard(df[y_pred.name],df[y_true.name]))
return scores
scores_by_year = scores_over_time(y_pred=linear_preds,y_true=y)
print(scores_by_year.tail(3).T)
scores_by_year['edge_to_mae'].plot(title='Prediction Edge vs. MAE')
Making the Ensemble Model¶
Now that we've trained base models and generated out-of-sample predictions, it's time to train the stacked generalization ensemble model.
Training the ensemble model simply requires feeding in the base models' predictions in as the X dataframe. To clean up the data and ensure the X and y are of compatible dimensions, I've created a short data preparation function.
Here, we'll use Lasso to train the ensemble becuase it is one of a few linear models which can be constrained to positive = True. This will ensure that the ensemble will assign either a positive or zero weight to each model, for reasons described above.
from sklearn.linear_model import LassoCV
def prepare_Xy(X_raw,y_raw):
''' Utility function to drop any samples without both valid X and y values'''
Xy = X_raw.join(y_raw).replace({np.inf:None,-np.inf:None}).dropna()
X = Xy.iloc[:,:-1]
y = Xy.iloc[:,-1]
return X,y
X_ens, y_ens = prepare_Xy(X_raw=pd.concat([linear_preds.rename('linear'),tree_preds.rename('tree')],
axis=1),y_raw=y)
ensemble_models,ensemble_preds = make_walkforward_model(X_ens,y_ens,algo=LassoCV(positive=True))
ensemble_preds = ensemble_preds.rename('ensemble')
print(ensemble_preds.dropna().head())
Note that the ensemble's predictions don't begin until July, since the earliest trained ensemble model isn't available until end of Q2. This is necessary to make sure the ensemble model is trained on out of sample data - and that its predictions are also out of sample.
Again, we can look at the coefficients over time of the ensemble model. Keep in mind that the coefficients of the ensemble model represents how much weight is being given to each base model. In this case, it appears that our tree model is much more useful relative to the linear model, though the linear model is gradually catching up.
pd.DataFrame([model.coef_ for model in ensemble_models],
columns=X_ens.columns,index=ensemble_models.index).plot(title='Weighting Coefficients for \nSimple Two-Model Ensemble')
Performance of Ensemble vs. Base Models¶
Now that we have predictions from both the base models and ensemble model, we can explore how the ensemble performs relative to base models. Is the whole really more than the sum of the parts?
# calculate scores for each model
score_ens = calc_scorecard(y_pred=ensemble_preds,y_true=y_ens).rename('Ensemble')
score_linear = calc_scorecard(y_pred=linear_preds,y_true=y_ens).rename('Linear')
score_tree = calc_scorecard(y_pred=tree_preds,y_true=y_ens).rename('Tree')
scores = pd.concat([score_linear,score_tree,score_ens],axis=1)
scores.loc['edge_to_noise'].plot.bar(color='grey',legend=True)
scores.loc['edge'].plot(color='green',legend=True)
scores.loc['noise'].plot(color='red',legend=True)
plt.show()
print(scores)
fig,[[ax1,ax2],[ax3,ax4]] = plt.subplots(2,2,figsize=(9,6))
metric = 'RSQ'
scores_over_time(y_pred=ensemble_preds.rename('ensemble'),y_true=y)[metric].rename('Ensemble').\
plot(title=f'{metric.upper()} over time',legend=True,ax=ax1)
scores_over_time(y_pred=linear_preds.rename('linear'),y_true=y)[metric].rename('Linear').\
plot(title=f'{metric.upper()} over time',legend=True, alpha = 0.5, linestyle='--',ax=ax1)
scores_over_time(y_pred=tree_preds.rename('tree'),y_true=y)[metric].rename("Tree").\
plot(title=f'{metric.upper()} over time',legend=True, alpha = 0.5, linestyle='--',ax=ax1)
metric = 'edge'
scores_over_time(y_pred=ensemble_preds.rename('ensemble'),y_true=y)[metric].rename('Ensemble').\
plot(title=f'{metric.upper()} over time',legend=True,ax=ax2)
scores_over_time(y_pred=linear_preds.rename('linear'),y_true=y)[metric].rename('Linear').\
plot(title=f'{metric.upper()} over time',legend=True, alpha = 0.5, linestyle='--',ax=ax2)
scores_over_time(y_pred=tree_preds.rename('tree'),y_true=y)[metric].rename("Tree").\
plot(title=f'{metric.upper()} over time',legend=True, alpha = 0.5, linestyle='--',ax=ax2)
metric = 'noise'
scores_over_time(y_pred=ensemble_preds.rename('ensemble'),y_true=y)[metric].rename('Ensemble').\
plot(title=f'{metric.upper()} over time',legend=True,ax=ax3)
scores_over_time(y_pred=linear_preds.rename('linear'),y_true=y)[metric].rename('Linear').\
plot(title=f'{metric.upper()} over time',legend=True, alpha = 0.5, linestyle='--',ax=ax3)
scores_over_time(y_pred=tree_preds.rename('tree'),y_true=y)[metric].rename("Tree").\
plot(title=f'{metric.upper()} over time',legend=True, alpha = 0.5, linestyle='--',ax=ax3)
metric = 'edge_to_noise'
scores_over_time(y_pred=ensemble_preds.rename('ensemble'),y_true=y)[metric].rename('Ensemble').\
plot(title=f'{metric.upper()} over time',legend=True,ax=ax4)
scores_over_time(y_pred=linear_preds.rename('linear'),y_true=y)[metric].rename('Linear').\
plot(title=f'{metric.upper()} over time',legend=True, alpha = 0.5, linestyle='--',ax=ax4)
scores_over_time(y_pred=tree_preds.rename('tree'),y_true=y)[metric].rename("Tree").\
plot(title=f'{metric.upper()} over time',legend=True, alpha = 0.5, linestyle='--',ax=ax4)
plt.tight_layout()
plt.show()
Observations:¶
- We can see that the ensemble is fairly consistently more effective than either of the base models.
- All models seem to be getting better over time (and as they have more data on which to train).
- The ensemble also appears to be a bit more consistent over time. Much like a diversified portfolio of stocks should be less volatile than the individual stocks within it, an ensemble of diverse models will often perform more consistently across time
Next Steps¶
The stacked generalization methodology is highly flexible. There are countless directions you can take this, including:
- More model types: add SVMs, deep learning models, regularlized regressions, and dimensionality reduction models to the mix
- More hyperparameter combinations: try multiple sets of hyperparameters on a particular algorithm.
- Orthogonal feature sets: try training base models on different subsets of features. Avoid the "curse of dimensionality" by limiting each base model to an appropriately small number of features.
Summary¶
This final post, combined with the five previous posts, presents an end-to-end framework for applying supervised machine learning techniques to financial time series data, in a way which helps mitigate the several unique challenges of this domain.
Please feel free to add to the comment section with your comments and questions on this post. I'm also interested in ideas for future posts.
Going forward, I plan to shift gears from tutorials to research on market anomalies and trading opportunities.
One more thing...¶
If you've found this post useful, please consider subscribing to the email list to be notified of future posts (email addresses will only be used for this purpose...).
You can also follow me on twitter and forward to a friend or colleague who may find this topic interesting.