Start here  ·  Coming from Excel  ·  The courses  ·  How a day works  ·  The data  ·  License

From Fixed Income to Python: A Practical Guide for Fixed Income Professionals

You know what a coupon is. You understand spread, duration, yield, and credit. You live in Excel. You understand the business, but you have never written a line of code.

That's exactly where this course begins.

Most Python courses teach programming through shopping carts, movie ratings, and toy datasets. This one teaches Python through the work of a credit desk.

Your first variables describe a bond: its ticker, its coupon, its maturity. Your first loop totals a portfolio. Your first data table is a holdings file. Your first analysis answers a question you already know how to ask.

Nothing is abstracted away into contrived examples. Every programming concept is introduced in the context of fixed income, portfolio management, and credit analysis, so you learn the code and immediately understand why it matters.

You already know the investment problem. This course teaches you how to solve more of it with code.

The series starts with Course 1, Python Foundations: four weeks, one notebook a day, 19 notebooks and a project. Course 2, Bond Math follows: three weeks, 14 notebooks and a capstone, in which you build price, yield, accrued interest, duration, DV01, convexity, spread, and DTS yourself, test every function against Excel, and finish with a risk report on a whole book. Course 3, Market Data and Time Series comes next: three weeks, 14 notebooks and a capstone, in which you pull market data from an API with a key that never leaks, clean it against the bond market's own calendar, compute changes, volatility, z-scores, percentile ranks, slopes, and butterflies with no lookahead, query the book with SQL, and finish with a daily market snapshot that runs again on any date. Course 4, Machine Learning follows: three weeks, 14 notebooks and a capstone, in which you fit and read regressions and classifiers on credit data, choose a threshold on rows the test set never saw, prune a decision tree, validate on dated data with no lookahead, cluster bonds and issuers, describe the Treasury curve with PCA, and finish with a three-model comparison on a public file of company statements. Course 5, Advanced Machine Learning comes next: three weeks, 14 notebooks and a capstone, in which you fit and read bagging, random forests, AdaBoost, gradient boosting, XGBoost, and LightGBM, stack models on out-of-fold outputs, tune without fooling yourself, engineer features from financial ratios, handle a rare outcome with weights, resampling, or a threshold, describe a model with SHAP values, and finish with a four-model comparison on a public file of company ratios. Course 6, Neural Networks follows: three weeks, 14 notebooks and a capstone, in which you compute gradients by hand and check them against TensorFlow, show that one neuron is a logistic regression, train networks in Keras one setting at a time, score recurrent and convolutional networks on the Treasury curve against carrying the last value forward, and finish with a network measured against LightGBM on a public file of company ratios. Course 7, NLP, RAG, and LLMs comes next: three weeks, 14 notebooks and a capstone, in which you read FOMC statements, 8-K filings, and real indentures with Python, turn text into TF-IDF vectors, word vectors, and sentence embeddings, classify and search it on the CPU, send requests to large language models with the key kept secret, extract an indenture's covenants with every number checked against the text, and finish with a covenant extraction tool run on an indenture you have never seen. Course 8, Implementing Agentic AI follows: three weeks, 14 notebooks and a capstone, in which you send a model its first tool request by hand, write the agent loop yourself, hand an agent your own bond math, the synthetic book, the Treasury curve, FOMC statements, and a real indenture as tools, reject any figure that is not the value of the tool result it cites, serve tools through MCP, rebuild the loop in the OpenAI Agents SDK, split the work between an orchestrator and four specialists, grade the desk against answer keys written first, stop it with guardrails and a budget, and finish with a multi-agent credit desk whose every number is traced; every agent run was recorded once and replays without a key. Course 9, Deployment comes next: three weeks, 14 notebooks and a capstone, in which you push your code to GitHub after a scan that proves nothing private goes with it, review every change by its diff, run your tests on Linux on every push, put your own bond math behind a FastAPI service at a public URL, serve the Course 8 desk's brief with no key on the server, give it Streamlit and Gradio front ends and a Hugging Face Space, package it in Docker, log and monitor every request, measure a model's drift, and finish with the desk's brief as a public app on free accounts; every notebook runs offline, with the cloud steps shown as masked transcripts. Course 10, Implementing RAG follows: three weeks, 14 notebooks and a capstone, in which you ask a model five questions about a 10-K with no document and grade the answers against keys written first, load the filing as pages and items, chunk it with a recursive splitter you write, embed the chunks with pinned models and search them exactly with NumPy and FAISS, add BM25 and reciprocal rank fusion, rerank with a cross-encoder, tune one setting at a time on development questions, filter by metadata across four filings, demand a citation after every sentence, measure recall and MRR on a labelled set, put LLM judges beside a strict grader, refresh the index when a new filing arrives, serve the retriever behind FastAPI and price it, and finish with a cited question-answering tool over a folder of public filings; every generated answer was recorded once and replays without a key, and the local-model route runs on Colab's T4 GPU. The Bonus course, Build Your Own Coding Agent, closes the series: one notebook in which you build a coding agent from an empty file on OpenAI's Responses API, with three tools that read, list, and edit files in one fenced folder and none that runs code, check the bond math it writes against Excel, give it a chat page with your own name on it, and deploy it to a private Hugging Face Space of your own; every model request replays without a key. Every model describes its dataset; nothing is a forecast, a rating, or a lending rule, and work on a real filing is descriptive extraction only. All ten courses and the Bonus course are ready.

This course is educational content created in a personal capacity. Nothing here is investment advice or a recommendation to buy or sell any security. All portfolio data is synthetic.

This product uses the FRED® API but is not endorsed or certified by the Federal Reserve Bank of St. Louis.

If you know Excel, you already know how to think about this

Nothing in this course asks you to forget Excel. Wherever Excel has a feature for the job, the new idea is introduced next to it, with the formula written out, so you are translating something you know and not learning from nothing.

In Excel you would	In Python you will	Course	Where
Fill a formula down a column	Write the formula once, in a loop or on a whole array	1	Week 1, Day 1 and Week 2, Day 1
Wrap a number in TEXT and join it to a sentence with &	Write an f-string with a format code	1	Week 1, Day 1
Write a nested IF	Write if, elif, else, one test per line	1	Week 1, Day 1
See #N/A, #DIV/0!, or #NAME? in a cell	Read an error that names the line and the cause	1	Week 1, Day 3
Put an IF beside a total that shows OK or CHECK	Write an assert that stops the run when the total is wrong	1	Week 1, Day 3
Turn on AutoFilter and sort a sheet	Filter and sort a table by condition	1	Week 2, Day 2
Use SUMIF and COUNTIF	Filter, then sum or count	1	Week 2, Day 2 and Day 5
Use COUNTBLANK and Remove Duplicates	Count missing values and repeated rows in one line each	1	Week 2, Day 4
Use VLOOKUP or INDEX and MATCH	Merge two tables on a shared column	1	Week 2, Day 3 and Day 5
Build a pivot table	Group by a column and summarize	1	Week 2, Day 3 and Day 5
Open and save a workbook with several sheets	Read a sheet into a table, and write tables to the sheets of a new workbook	1	Week 2, Day 3 and Day 5
Use SUMPRODUCT over SUM for a weighted average	Do the same arithmetic on two columns	1	Week 2, Day 3 and Day 5
Insert a chart: Column, Line, Scatter, Histogram, Box and Whisker	Draw the same chart in a few lines of code	1	Week 3, Day 1
Add a trendline, and use SLOPE, INTERCEPT, and RSQ	Fit the line and print its numbers with units	1	Week 3, Day 2
Use Remove Duplicates, TRIM, and QUARTILE.INC to clean and check a sheet	Find and fix the same problems in code, and print the totals before and after	1	Week 3, Day 3
Use XLOOKUP by date, and AVERAGEIFS by year	Select by date and group a time series by year	1	Week 3, Day 4
Net buys against sells with two SUMIFS, and keep a running total	Sign the sells, group, and take a cumulative sum	1	Week 3, Day 5
Use AVERAGEIF on a column of 0s and 1s to get a rate by group	Group, then take the mean and the count together	1	Week 4, Day 1
Use FIND, SUBSTITUTE, TRIM, and Find and Replace on text	Search and clean text with patterns, across every row at once	1	Week 4, Day 2
Count words from a list with SUMPRODUCT, and standardize with STANDARDIZE	Score every document against a word list and compare the scores	1	Week 4, Day 3
Paste next month's rows into last month's workbook and re-point every range	Keep the steps as functions in one file and run them on the new file	1	Week 4, Day 4
Use FV and PV, and read the minus sign PV returns	Write future_value and present_value and test them against Excel's values	2	Week 1, Day 1
Price a bond on a coupon date with NPV over a column of cash flows, -PV with a payment, or PRICE	List the cash flows and discount them in one function	2	Week 1, Day 2
Fill a column of PRICE formulas down a range of yields and chart it as a Scatter with Straight Lines	Price a grid of yields and plot the price-yield curve	2	Week 1, Day 3
Restate a yield with EFFECT and NOMINAL	Write convert_yield between any two compounding frequencies	2	Week 1, Day 3
Goal Seek a yield, halve a bracket with IF formulas, or use RATE and YIELD	Write a bisection loop that solves the yield from a price	2	Week 1, Day 4
Round a coupon with MROUND, and open a workbook of live PRICE and YIELD formulas	Find the coupon on a 1/8 grid, and write a workbook of live formulas with openpyxl	2	Week 1, Day 5
Read date serial numbers, and use EDATE, EOMONTH, COUPPCD, COUPNCD, and COUPNUM	Work with date values, add months, and build a coupon schedule back from maturity	2	Week 2, Day 1
Use COUPDAYBS, COUPDAYS, COUPDAYSNC, ACCRINT, and YEARFRAC beside DAYS360	Count days on US 30/360 and actual/actual, and compute accrued interest	2	Week 2, Day 2
Use PRICE between coupon dates, add accrued for the dirty price, use YIELD, and write the price out cell by cell	Write price, dirty_price, and yield_to_maturity, and invoice a month of trades	2	Week 2, Day 3
Use DURATION and MDURATION, build DV01 from cells, and bump the yield in two PRICE cells	Write duration and DV01 functions, and check DV01 by bumping the yield	2	Week 2, Day 4
Estimate convexity from three PRICE cells, and total a book's DV01 with SUMPRODUCT	Write convexity, and reprice a whole book with its DV01 and duration	2	Week 2, Day 5
Interpolate a curve with FORECAST.LINEAR, INDEX, and MATCH, and chart it as a Scatter	Write interpolate_yield, with flat ends, and compare it with numpy.interp	2	Week 3, Day 1
Use YEARFRAC for years to maturity, and subtract the curve yield for a spread	Measure every bond's spread to the curve, and reproduce the file's spread column	2	Week 3, Day 2
Divide two bumped PRICE cells for spread duration, multiply for DTS, and total buckets with SUMIFS	Write spread_duration and dts, and group each bond's share of the book's DTS	2	Week 3, Day 3
Add a calculated field to a pivot table, and run a What-If Data Table over a whole-book reprice	Group by sector and rating, and reprice the book under rate and spread scenarios	2	Week 3, Day 4
Check a risk report workbook with SUM and SUMPRODUCT, and price a month-end and a final-period bond with the COUP functions, PRICE, and YIELD	Clean, recompute, and reconcile a month-end extract, and write the risk report to Excel	2	Week 3, Day 5
Build a request URL in cells with & and ENCODEURL, and load a web CSV with Power Query From Web	Build the request in code, and read the key from .env so it is never typed into a cell or a file	3	Week 1, Day 1
Load JSON with Power Query, and turn FRED's . into #N/A with IFERROR(VALUE(...),NA())	Parse FRED's JSON, reading . as missing and never as 0	3	Week 1, Day 2
Use WEEKDAY, WORKDAY, and NETWORKDAYS with a holiday list, and see XLOOKUP return #N/A on a date with no row	Take the market calendar from the data's own dates, and join two series inner or outer	3	Week 1, Day 3
Find a month's last market date with EOMONTH and MAXIFS, look up a value as of a date with XLOOKUP match mode -1, and group a pivot table by month	Take the last observation of each week or month, labelled by its own date, and the as-of value	3	Week 1, Day 4
Append and merge files with Power Query, and check a sheet with SUMPRODUCT, UNIQUE, and COUNTBLANK	Build one checked daily dataset from two files, and cache a download	3	Week 1, Day 5
Write =(B3-B2)*100 for a change in bp, XLOOKUP with EDATE for a one-month change, and MAX with XLOOKUP for the date of the largest move	Compute changes in bp over rows and over calendar periods, and price returns in percent	3	Week 2, Day 1
Fill STDEV.S down a fixed-size range, with STDEV.P, OFFSET, and *SQRT(252) beside it	Compute rolling volatility over a trailing window of rows, and test that it uses no future data	3	Week 2, Day 2
Use STANDARDIZE, and COUNTIF over COUNT for a percentile rank, beside PERCENTRANK.INC, RANK.EQ, and RANK.AVG	Compute rolling z-scores and percentile ranks that match Excel row for row	3	Week 2, Day 3
Fill slope and butterfly formulas down the Treasury sheet, and weight a butterfly with PRICE and MDURATION	Compute slopes and butterflies in bp, and DV01-neutral weights from your own bondmath	3	Week 2, Day 4
Build a summary sheet for one date with XMATCH, XLOOKUP, INDEX, OFFSET, STANDARDIZE, and COUNTIF	Build every measure as of one date in one table that reads nothing after it	3	Week 2, Day 5
Use COUNTIFS, SUMIFS, AVERAGEIFS, FILTER, TAKE with SORTBY, and a pivot table	Write WHERE, GROUP BY, HAVING, ORDER BY, and LIMIT in SQL, each checked against pandas	3	Week 3, Day 1
Join with XLOOKUP or Power Query Merge, and find repeated keys with COUNTIF	Join tables in SQL, catch a join that duplicates rows, and use window functions for daily changes and month ends	3	Week 3, Day 2
Check a suggested formula, PERCENTRANK.INC, against COUNTIF over COUNT, and a DV01 built from cells	Test code written in the style of an AI assistant against your own functions and Excel's values	3	Week 3, Day 3
Open the snapshot workbook and check its numbers with SUMIFS, STANDARDIZE, and COUNTIF	Write a daily market snapshot to HTML and Excel from one function that runs on any date	3	Week 3, Day 4
Test a pasted extract's dates and units with DATE, WEEKDAY, COUNTIF, and IFERROR(VALUE(...),NA())	Clean a vendor extract with checks that stop, reconcile it, and run the snapshot on two dates	3	Week 3, Day 5
Fit SLOPE, INTERCEPT, and RSQ on a training sheet, and score a test sheet with AVERAGE(ABS(...)) and SQRT(SUMXMY2(...)/COUNT(...))	Split the rows, fit on the training rows, and measure the error in bp on rows the model never saw	4	Week 1, Day 1
Build dummy columns with =--(D2="AAA"), fit them with LINEST, and score the test rows with TREND	Encode ratings with a named reference level and fit a multiple regression	4	Week 1, Day 2
Read LINEST's standard errors, and compute p-values with T.DIST.2T, intervals with T.INV.2T, a VIF with 1/(1-R2), and a joint F-test with F.DIST.RT	Read statsmodels' coefficient table, test sector jointly, and measure multicollinearity	4	Week 1, Day 3
Take LN and EXP of spreads, run Breusch-Pagan by hand with LINEST and CHISQ.DIST.RT, and compute Q-Q points with NORM.S.INV	Check a regression's assumptions with residual plots and tests, and model log spreads	4	Week 1, Day 4
Standardize with STANDARDIZE and STDEV.P, and solve ridge in closed form with MMULT, MINVERSE, and MUNIT	Fit ridge and lasso inside a pipeline that scales on the training rows only	4	Week 1, Day 5
Fill a logistic probability formula, sum a log-likelihood column, and maximise it with Solver, the non-negative box turned off	Fit a logistic regression and read odds ratios with exp	4	Week 2, Day 1
Count a confusion matrix with COUNTIFS against a threshold cell	Count the confusion matrix and compute precision, recall, and F1 for the default class	4	Week 2, Day 2
List thresholds with SORT(UNIQUE(...)), trace the ROC with COUNTIFS, take the AUC with SUMPRODUCT, and pick a threshold with MAXIFS	Draw ROC and precision-recall curves, and choose a threshold for a target recall on validation rows	4	Week 2, Day 3
Compute Gini with COUNTIFS, a leaf's default share with AVERAGEIFS, and write a small tree as nested IF	Grow, prune, and read a decision tree	4	Week 2, Day 4
Open the results workbook, recompute precision and recall, and find the best model with INDEX and MATCH	Compare three models on validation rows, open the test rows once, and write the comparison to Excel	4	Week 2, Day 5
Fill SLOPE(OFFSET(...)) down a trailing window, fit one fold with TREND, and compare a whole-history STANDARDIZE with a trailing one	Validate on dated rows with walk-forward splits, and show lookahead in a feature	4	Week 3, Day 1
Measure distance to each centre with SQRT(SUMXMY2(...)), assign with MATCH(MIN(...)), and update the centres with AVERAGEIF	Scale the bonds and run K-means, with the elbow and the silhouette	4	Week 3, Day 2
Count clusters with a pivot table, and take medians with MEDIAN(IF(...)), which a pivot table cannot give	Build a hierarchical clustering, cut it, and profile the clusters in their own units	4	Week 3, Day 3
Build a COVARIANCE.S grid, and score each date with MMULT of the centred changes	Find level, slope, and curvature with PCA on the Treasury curve	4	Week 3, Day 4
Check the capstone's results workbook with COUNTIFS against its threshold	Compare three models on a new file, test once, and write the results to Excel	4	Week 3, Day 5
Chart a comparison of models as a clustered column or line chart, and shade a table with a conditional formatting color scale	Draw every comparison as a chart and show it as a formatted table with the best value in each column marked	4	Week 1, Day 1 to Week 3, Day 4
Compute the share a sample leaves out with =(1-1/n)^n, count pasted positions with FREQUENCY and COUNTIF, and see RANDARRAY redraw on every recalculation	Draw a seeded bootstrap, count its rows, and measure how ten trees on ten samples disagree	5	Week 1, Day 1
Average pasted tree columns with AVERAGE, and score a vote with SUMPRODUCT	Bag trees by hand and with scikit-learn, and read the out-of-bag score as the accuracy it is	5	Week 1, Day 2
Count columns per split with INT(SQRT(19)), the chance one is tried with COMBIN, and two trees' agreement with CORREL	Fit a random forest, measure tree correlation, and summarise a search's fold scores	5	Week 1, Day 3
Rank two importance columns with RANK.EQ, and compute a ROC AUC from RANK.AVG and SUMIF	Compare impurity and permutation importance, and catch a column of noise that one of them ranks second	5	Week 1, Day 4
Turn a fitted log spread into bp with EXP, average the misses with AVERAGE(ABS(...)), and compute R-squared with SUMXMY2 and DEVSQ	Score OLS, a tree, and a forest on the spread cross-section by cross-validation and by issuer	5	Week 1, Day 5
Run an AdaBoost round with SUMPRODUCT, LN, and EXP, and see LOG return base 10	Run three AdaBoost rounds by hand and match scikit-learn's stump weights	5	Week 2, Day 1
Fit one boosting round from residuals with two AVERAGEIFS and an IF	Build gradient boosting from residuals and match the library to 1e-12	5	Week 2, Day 2
Add gradients with SUMIFS for an XGBoost leaf and its gain, and fill a log loss column with LN	Regularise XGBoost, stop it early on rows carved from the training rows, and compute log loss by hand	5	Week 2, Day 3
Blend two pasted probability columns with AVERAGE, and apply a stack's logistic layer with EXP	Stack models on out-of-fold outputs and check whether the stack beats its best member	5	Week 2, Day 4
Read a bootstrap interval with PERCENTILE.INC, and find the chosen model with INDEX and MATCH	Compare six models with bootstrap intervals and paired differences, and write the results to Excel	5	Week 2, Day 5
Summarise five fold scores with AVERAGE and STDEV.P, not STDEV.S	Read a randomized search's results and report a nested score beside its best score	5	Week 3, Day 1
Cap a column with MIN, guard a ratio with IF(OR(...),NA(),...), clip with PERCENTILE.INC, and count gaps with COUNTBLANK	Engineer features from financial ratios, each statistic learned inside the fold	5	Week 3, Day 2
Rebuild one SMOTE row with =x+u*(nb-x), correct a resampled score by formula, and compute a Brier score with SUMXMY2	Compare class weights, resampling, and a threshold on a rare outcome, and recalibrate the score	5	Week 3, Day 3
Add a SHAP row with SUM to the model's log-odds, turn it into a probability with EXP, and average absolute values with AVERAGE(ABS(...))	Split a company's score across its ratios with SHAP values, and never read a contribution as a cause	5	Week 3, Day 4
Check the capstone's confusion matrix with COUNTIFS, its interval with PERCENTILE.INC, and a SHAP row with SUM	Compare four models on a new file, test once, describe the chosen one with SHAP values, and write the results to Excel	5	Week 3, Day 5
Chart a comparison of models as a clustered column, bar, line, or scatter chart, add custom error bars for a spread of fold scores, and shade a table with a conditional formatting color scale	Draw every comparison as a chart and show it as a formatted table with the best value in each column marked	5	Week 1, Day 1 to Week 3, Day 4
Fill a neuron's output =1/(1+EXP(-($H$1+$H$2*A2))) down a column, take its gradient with SUMPRODUCT over ROWS, and fill a 100-step descent path	Compute a log loss and its gradient by hand, check them against tf.GradientTape, and take gradient steps	6	Week 1, Day 1
Maximise a log-likelihood with Solver, scale columns with AVERAGE and STDEV.P, and score a row with SUMPRODUCT and EXP	Fit one Keras neuron and show that it is a logistic regression, with the same odds ratios	6	Week 1, Day 2
Build a hidden layer from SUMPRODUCT, MAX, and TANH, and count a network's weights with =19*32+32+32*1+1	Add hidden layers, reproduce a network's output in numpy, and count its parameters	6	Week 1, Day 3
Count steps per epoch with ROUNDUP, and find the best epoch on a pasted loss column with MATCH(MIN(...))	Read loss curves, change the batch size, and stop early on rows carved from the training rows	6	Week 1, Day 4
Count a confusion matrix with COUNTIFS on a flag column, and read a bootstrap interval with PERCENTILE.INC	Put a network beside logistic regression and LightGBM, with thresholds from out-of-fold outputs and paired intervals	6	Week 1, Day 5
Compute MSE with SUMXMY2, MAE with AVERAGE(ABS(...)), a Huber row with IF, and standardize a column with STDEV.P	Fit regression networks with three losses, turn the target back to bp, and compare them with OLS by five folds	6	Week 2, Day 1
Fill a momentum path and three Adam steps with the bias correction, and a decaying learning rate with =$B$1*$B$2^(A6/$B$3)	Check momentum and Adam steps against Keras, and choose an optimizer and a learning rate on the stopping rows	6	Week 2, Day 2
Normalize a batch with AVERAGE, VAR.P, and SQRT, and compute Glorot's limit and He's spread with SQRT	Start a network from the right weights, see gradients vanish, and add batch normalization	6	Week 2, Day 3
Drop values with =A2*(B2>=0.2)/(1-0.2), and compute an L2 penalty with SUMSQ	Brake overfitting with L2, dropout, and early stopping, scored by five folds	6	Week 2, Day 4
Compute balanced class weights with ROWS and COUNTIF, and a weighted log loss with SUMPRODUCT over SUM	Fit a class-weighted network, read its output as a score, and decide on weights by numbers	6	Week 2, Day 5
Write =(B3-B2)*100 for a change in bp, average a 20-day window with OFFSET, and score zero change with AVERAGE(ABS(...))	Build windows of past changes with no lookahead, and score a network walk-forward against carrying the last value forward	6	Week 3, Day 1
Take a SimpleRNN and an LSTM step with TANH, EXP, and MMULT, and compare two spreads with STDEV.P	Fit SimpleRNN, LSTM, and GRU networks walk-forward, and check one LSTM step against Keras	6	Week 3, Day 2
Fill a weighted moving sum =SUMPRODUCT(B2:B6,$K$2:$K$6)+$K$7 down a column, then pool with MAX and AVERAGE	Fit a one-dimensional convolution, feed all 11 tenors in as channels, and keep same-day changes out	6	Week 3, Day 3
Average errors by fold with AVERAGEIFS, and count the folds a rule wins with COUNTIF over COUNT	Choose a sequence model on the development years, open the held-out years once, and write the results to Excel	6	Week 3, Day 4
Check the capstone's workbook with AVERAGE, STDEV.P, paired differences, COUNTIFS, and PERCENTILE.INC	Compare a network with LightGBM by paired folds on three fold seeds, with intervals, and write the results to Excel	6	Week 3, Day 5
Chart a loss curve or a comparison of networks as a line, clustered column, bar, or histogram chart, set a logarithmic axis, and shade a grid with a conditional formatting color scale	Draw every loss curve and every comparison as a chart and show it as a formatted table with the best value in each column marked	6	Week 1, Day 1 to Week 3, Day 4
Find a word with SEARCH and FIND, count it with LEN and SUBSTITUTE (a substring count), split a sentence with TEXTSPLIT, and pull the target range with REGEXEXTRACT	Normalize look-alike characters, pull numbers out of text with regular expressions, and split text into tokens	7	Week 1, Day 1
Count the statements that hold a word with COUNTIF(range,"*inflation*"), compute an IDF with LN, scale a row with SQRT(SUMSQ(...)), and a cosine with SUMPRODUCT	Build TF-IDF by hand, check it against scikit-learn, and compare documents by cosine similarity	7	Week 1, Day 2
Count a confusion matrix with COUNTIFS on item codes stored as text, and see COUNTIF count "2.1" as "2.10" where EXACT does not	Train a TF-IDF classifier on 8-K sections split by date, catch the heading leak, and report recall on the rare items	7	Week 1, Day 3
Compute the cosine of two pasted word vectors with SUMPRODUCT and SUMSQ, and see MATCH and XLOOKUP ignore case where EXACT does not	Find a word's nearest neighbours in pretrained vectors, and use mean vectors as classifier features	7	Week 1, Day 4
Count a statement's words from two named lists with SUMPRODUCT, LEN, and SUBSTITUTE, and score them per thousand words	Score every statement with a word list and with a classifier fitted only on earlier years, the label's own sentence removed first	7	Week 1, Day 5
Compute a question's score against a chunk as SUMPRODUCT over 384 pasted numbers, and check each vector's length with SQRT(SUMSQ(...))	Embed an indenture's chunks and search it by meaning	7	Week 2, Day 1
Read the item over a row's largest score with INDEX, MATCH, and MAX, and the share where two models agree with SUMPRODUCT	Score a zero-shot classifier and an embedding classifier against TF-IDF, reading every score by its label's name	7	Week 2, Day 2
Count risk factors by cluster and filing with COUNTIFS	Cluster risk factor embeddings with K-means and describe each cluster by its words and its counts	7	Week 2, Day 3
Cut an item out of pasted filing text with TEXTBEFORE and TEXTAFTER, and count its words with TEXTSPLIT	Turn an EDGAR filing into text and cut a 10-K into its items	7	Week 2, Day 4
Check three models' confusion matrices on the test rows with COUNTIFS, then recall, precision, F1, and AVERAGE	Open the test sections once and compare three models with paired bootstrap intervals	7	Week 2, Day 5
Price a request with =B2*C2+B3*C3 on token counts and per-token prices, and see SUMPRODUCT treat prices pasted as text as 0	Send a request to an LLM from code with the key kept secret, count its tokens, and price it	7	Week 3, Day 1
Check a quote against its section with =LEN(C2)<>LEN(SUBSTITUTE(C2,B2,"")), because SEARCH fails on long quotes and reads * and ? as wildcards	Extract a covenant into a schema and check every quote and number against the indenture	7	Week 3, Day 2
Pull a grower basket out of a pasted sentence with TEXTBEFORE(TEXTAFTER(A2,"greater of "),"%"), and see a cell cut a long section at 32,767 characters	Check a model's baskets against a regular expression, put two indentures in one table, and read a PDF page as text and as an image	7	Week 3, Day 3
Score retrieval with ISNUMBER(MATCH(B2,C2:G2,0)), with the section numbers kept as text	Retrieve an indenture's chunks for a question, answer with citations, and judge each answer against the retrieved text	7	Week 3, Day 4
Check the capstone's workbook, whose quote and number checks are SUBSTITUTE and SEARCH formulas	Build a covenant extraction tool that runs on an indenture you have never seen, with a checked table, and write it to Excel	7	Week 3, Day 5
Chart a count, a score, or a comparison of models as a line, clustered column, stacked column, bar, XY scatter, or histogram chart, and shade a COUNTIFS grid or a cosine table with a conditional formatting color scale	Draw every count, score, embedding, and comparison as a chart and show each comparison as a formatted table with the best value in each column marked	7	Week 1, Day 1 to Week 3, Day 4
Build a DV01 from PRICE, MDURATION, and the accrued from COUPDAYBS over COUPDAYS, and price a run with =B2*C2+B3*C3, where SUMPRODUCT treats prices pasted as text as 0	Hand a model your own dv01 as a tool, run the tool it requests on your machine, and price the run from its token counts	8	Week 1, Day 1
Count a pasted trace table's rows for a tool with COUNTIF(C2:C7,"price"), its requests with COUNTIFS, and its tokens with SUM	Write the agent loop with a step limit, and read every run as a trace table before its answer	8	Week 1, Day 2
Use PRICE, YIELD, MDURATION, and COUPDAYBS over COUPDAYS on one bond, and build its DV01 from three cells	Offer Course 2's bond functions to a model as tools, each docstring stating its units, and check the tool results against Excel	8	Week 1, Day 3
Check a figure against its tool value with =ROUND(B2,D2)=C2, at the decimals the answer wrote	Ask for an answer in a schema, and reject any figure that is not the value of the tool result it cites	8	Week 1, Day 4
Total a sector with SUMIFS over a helper column, take the k-th largest with LARGE, weight by market value with SUMPRODUCT over SUM, and find a CUSIP with INDEX and MATCH	Compute every answer key with pandas first, then check a risk agent's figures on the synthetic book against them	8	Week 1, Day 5
Look up a yield and a change in bp with INDEX and MATCH(...,0), and see #N/A for a date the file lacks	Give an agent the Treasury curve as tools that return an error for a missing date, never the nearest row	8	Week 2, Day 1
Check a cited section with =ISNUMBER(MATCH(B2,C2:G2,0)), the section numbers kept as text	Give an agent documents as tools, and check every section its answer cites against the sections the run read	8	Week 2, Day 2
Compute a change and a slope in bp as the difference of two INDEX and MATCH cells times 100	Serve your market tools from an MCP server, and send the model exactly the requests it saw before	8	Week 2, Day 3
Write each curve change with INDEX, MATCH, and *100, and see the change in percent when *100 is missing	Run a macro agent on the last two FOMC statements, and check its note's figures against pandas and its quotes against the statements	8	Week 2, Day 5
Total requests and tokens by agent with SUMIF, which matches "pm" to PM where pandas' == does not	Split the work between an orchestrator and four specialists, and keep the run log by agent	8	Week 3, Day 1
Look up a value by key and date with =XLOOKUP(1,(A2:A4="10 Yr par yield")*(D2:D4="2026-10-02"),B2:B4,"no value for that date")	Keep a desk memory that returns a value only for the date asked, never the latest one	8	Week 3, Day 2
Count passed rows by kind with COUNTIFS, and average requests with AVERAGEIFS, on the eval log	Grade an agent on questions whose answer keys were decided first, and keep the scorecard in Git	8	Week 3, Day 3
Check a request budget with =IF(SUM(B2:B10)>=10,"stop","ok"), where > lets one more request through	Stop a run in code, with guardrails and a budget checked before each request is sent	8	Week 3, Day 4
Check the capstone's workbook, whose figure checks are =ROUND(G2,H2)=C2 and whose scorecard is COUNTIFS and AVERAGEIFS	Build a multi-agent credit desk with every figure traced and the guardrail asserted, and write its checks to Excel	8	Week 3, Day 5
Total a size table in MB with =SUM(C2:C200)/1024^2, and judge it against a 500 MB limit with =IF(E1>500,"over","fits")	List what a server must install from a file's imports, measure it with its dependencies, and see why the desk ships as a snapshot	9	Week 1, Day 4
Count green runs on main with =COUNTIFS(D2:D9,"success",B2:B9,"main"), and take a pass rate that leaves cancelled runs out	Read a repo's CI runs as a table and a pass rate, and follow one change through a red run and a green one	9	Week 1, Day 5
Price a bond with PRICE, take the accrued from COUPDAYBS over COUPDAYS and the DV01 from MDURATION times the dirty price over 10,000, and read a service's /health with WEBSERVICE, which sends GET only	Put bondmath behind a FastAPI service, test it without a network, and match Excel at 6 decimals on six bonds	9	Week 2, Day 1
Read the deployed DV01 into a cell with WEBSERVICE and =VALUE(TEXTBEFORE(TEXTAFTER(A2,"""dv01_per_100"":"),"}")), equal at Excel's 15 significant digits, and see a refused request come back as #VALUE!	Deploy the service to a public URL on Vercel and check it against bondmath on your own machine	9	Week 2, Day 2
Price the desk's saved run with =SUMPRODUCT(B2:B3,C2:C3) on tokens and prices per token, where prices typed into text cells count as 0	Serve the Course 8 desk's brief as a snapshot with no key on the server, and refuse a request for a view at the API	9	Week 2, Day 3
Open a CSV from a download button and see Excel turn the CUSIP 99123E105 into the number 9.9123E+109, keep the course's CUSIPs as text, and read True and False as booleans for COUNTIF	Give the same functions a Streamlit page and a Gradio app, and offer identifiers in a workbook, where they stay text	9	Week 2, Day 4
Summarise a request log with COUNTIFS and AVERAGEIFS, take p95 latency with PERCENTILE.INC over FILTER, and see PERCENTILE.EXC and SMALL with INT give other answers on 20 values	Log every request without its body, and read requests, errors, and p50 and p95 latency by path, with alerts by rule	9	Week 3, Day 2
Score drift with =STANDARDIZE(AVERAGE(recent),AVERAGE(train),STDEV.S(train)), not STDEV.P, and count inputs outside the training range with two COUNTIFS against MAX and MIN	Keep a model as a file with its card, serve it, and measure how far its inputs have moved from its training window	9	Week 3, Day 3
Cost a live month with =SUMPRODUCT(C2:C5,D2:D5) on tokens and prices, and mark each free-tier line with =IF(B2>C2,"over","within")	Keep secrets in platform settings, list every name the code reads, and meter a month against the free tiers	9	Week 3, Day 4
Check the capstone's workbook, whose free-tier status is IF and whose model cost is SUMPRODUCT	Ship the desk's brief as a public app with CI, a request log, a daily cap, and a cost line, and write its numbers to Excel	9	Week 3, Day 5
Grade an answer against its key with ISNUMBER(SEARCH(key,answer)) and count the matches by stage with COUNTIFS	Grade a model's closed-book and prompt-only answers against keys written first, and see that none matches	10	Week 1, Day 1
Paste a page one line per row and read LEN of each, and see a line such as December 31, 2025 become a date serial in a General cell	Load a 10-K as pages with pdf_pages, keep the empty page, and map every page to its item	10	Week 1, Day 2
Count a chunk's characters with LEN and its words with LEN and SUBSTITUTE (with CHAR(10) turned into a space), and check an overlap with =RIGHT(B2,64)=LEFT(B3,64)	Write a recursive splitter with size and overlap, and read what 64 characters carry across a cut	10	Week 1, Day 3
Compute a cosine with SUMPRODUCT over SQRT(SUMSQ()) on two pasted 384-number vectors	Embed the chunks with three pinned models and search them exactly with NumPy and FAISS	10	Week 1, Day 4
Count matched keys by stage with COUNTIFS, and fill a 2 by 2 table of retrieval against grade with three COUNTIFS	Run the first RAG pipeline on five questions beside the closed-book and prompt-only stages, and say which miss was retrieval	10	Week 1, Day 5
Write one BM25 term weight out with LN and the length penalty, and fuse two rankings with =1/(60+rank) summed by SUMIFS	Write BM25 by hand, check it against rank-bm25, and fuse it with embedding search by reciprocal rank fusion	10	Week 2, Day 1
Read a tuning grid with AVERAGEIFS and order it with SORTBY	Run twelve named settings on development questions and read the results table	10	Week 2, Day 3
Filter a corpus table with COUNTIFS, ROWS(FILTER()), and TAKE(SORTBY(FILTER())), and see 4.10 typed into a General cell become 4.1	Filter chunks by document and section before search, across four filings	10	Week 2, Day 4
Check each cited chunk against the chunks sent with ISNUMBER(MATCH()) on ids stored as text	Demand a citation after every sentence and reject the answer that has none	10	Week 2, Day 5
Compute recall with ISNUMBER(MATCH()) and COUNTIF over ROWS, and MRR with AVERAGE(IFERROR(1/MATCH(),0))	Measure retrieval recall at k and MRR on a labelled question set, by item	10	Week 3, Day 1
Fill a 3 by 3 agreement table with COUNTIFS, write Cohen's kappa out as (po-pe)/(1-pe), and colour the table with a colour scale	Put two LLM judges beside a strict grader and read the agreement table and kappa	10	Week 3, Day 2
Match chunk texts with EXACT inside SUMPRODUCT, where COUNTIF and MATCH refuse a criterion over 255 characters, and see MAXIFS return 0 for a date before the first filing	Refresh an index when a new filing arrives, embedding only the new chunks, and answer as of each filing's date	10	Week 3, Day 3
Price a run with SUMPRODUCT on tokens and prices per token, where prices pasted as text count as 0, and average by setting with AVERAGEIFS	Serve the retriever behind a FastAPI service and price a thousand answers	10	Week 3, Day 4
Check the capstone's workbook, whose recall, MRR, COUNTIFS grids, kappa, and SUMPRODUCT bill are formulas	Build a cited question-answering tool over public filings with its evaluation report, and write its numbers to Excel	10	Week 3, Day 5
Open bonds.csv and see bond_id stay text and the maturity become a date, shown as ######## until the column is widened, then list the rows below BBB- with =FILTER(A2:F7,ISNUMBER(MATCH(E2:E7,{"BB+","BB","BB-","B+","B","B-"},0)))	Ask your own coding agent which bonds in the file are high yield, and check its answer against the file	Bonus	Week 1, Day 1
Price with =PV(0.06,3,-5,-100) and =PV(0.05,3,-5,-100), and take =MDURATION(DATE(2026,10,2),DATE(2029,10,2),0.05,0.06,1)	Have the agent write bond_price and modified_duration, and check them against 97.3270, 100.0, and 2.6956 before you trust them	Bonus	Week 1, Day 1
Count an agent's edits with =COUNTIFS(B2:B40,"edit_file") and its refused requests with =COUNTIFS(D2:D40,"ERROR*"), and price its run with =SUMPRODUCT(B2:B40*$G$1)+SUMPRODUCT(C2:C40*$G$2)	Read the agent's tool log as a table, and price its saved run from the token counts	Bonus	Week 1, Day 1

What changes is what you get to keep. A spreadsheet holds the answer. Code holds the steps that produced it, so the same steps can be run again on tomorrow's file without anyone dragging a formula or repointing a range.

Start here

There are two ways to run the notebooks. Pick one.

Run in Google Colab, with nothing to install

Open any notebook in the Course 1 map and click the Open in Colab badge at the top. Colab runs in your browser and needs only a Google account.

Use Colab on a personal account, with the course data only. Do not upload anything from your employer.

Run on your own Windows machine

You need Python 3.13, VS Code with the Python and Jupyter extensions, and git. On a work machine, check your firm's policy before installing anything.

Two things to get right before Step 1.

Check	What to do
Python on the PATH	In the Python installer, tick Add python.exe to PATH on the first screen. If python is still not found later, use py -3.13 wherever a step says python.
Scripts allowed in PowerShell	Run Set-ExecutionPolicy -Scope CurrentUser RemoteSigned once. Without it, Step 2 stops at Activate.ps1 with a message that running scripts is disabled.

Step 1. Get the course and open its folder.

mkdir $HOME\projects -Force
cd $HOME\projects
git clone https://github.com/lenamonj/python-fixed-income.git
cd python-fixed-income

No git, or not allowed to install it? On the repo's GitHub page choose Code, then Download ZIP, unzip it into projects, and cd into the unzipped folder.

Step 2. Create an environment and install everything the series needs, in one step. The root requirements.txt pins every library of Courses 1 to 10 and the Bonus course.

python -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt

It is a large install: TensorFlow and PyTorch are in it, so allow time and a few gigabytes of disk. To install one course only, use that course folder's own requirements.txt instead; the files are cumulative, so a later course's file covers the earlier ones.

Every notebook also opens with an install section: its first code cell checks the course's pinned libraries and installs only what is missing, so in Google Colab there is nothing to set up, and on a machine prepared with Step 2 it prints that everything is present.

What the courses need beyond the install:

• Courses 1, 2, 4, 5, 6, 9 and 10 need no API key. Course 2 uses git from its first day.
• Course 3 uses a free FRED API key of your own, kept in a .env file outside the course repo or in Colab's Secrets panel, never in the repo; every notebook runs without one. The Course 3 README explains.
• Course 7's week 3 asks for a free OpenRouter key of your own (an OpenAI key works as the alternative route), kept in my_bondmath\.env or Colab's Secrets panel; with no key every week 3 notebook replays the course's saved responses. The Course 7 README explains. Its two small models download into the Hugging Face cache, about 118 MB, the first time a notebook needs them.
• Course 8 needs no key: every agent run was recorded once and replays from the repo; an OpenAI or OpenRouter key of your own lets a notebook go live. The Course 8 README explains.
• Course 9 deploys to your own free GitHub, Vercel, and Hugging Face accounts and runs Docker Desktop on your own machine; create the Hugging Face account before you start, because a free account hosts a Space only once it is older than 30 days. Every notebook runs offline. The Course 9 README explains.
• Course 10 replays every generated answer and judge verdict from its saved runs; a key of your own in my_bondmath\.env makes the same cells live. In Colab it advises the T4 GPU runtime for its local-model cells and never requires it. The Course 10 README explains.
• The Bonus course replays its saved responses; an OpenAI key of your own in my_bondmath\.env makes the notebook live, and its last step deploys the agent to a private Space on the Hugging Face account you created for Course 9. The Bonus course README explains.

Step 3. Open the folder in VS Code.

code .

Open course1_python_foundations\week1\day1\01_python_for_fixed_income_intro.ipynb, choose the .venv kernel in the top right, and run the first cell.

Two Course 1 notebooks, Week 4 Days 2 and 3, download two small word lists the first time they run, so they need an internet connection once.

How a day works

Inside a course, each day has its own folder, and each folder holds three files with the same name.

File	What it is
NN_name.ipynb	The notebook. Work through it top to bottom.
NN_name_overview.pdf	A short slide overview of the day's ideas. Read it first.
NN_name_solutions.ipynb	Worked answers to the exercises. Try each one yourself before you look.

Every notebook teaches the same way.

Step	What happens
A question	A question the desk would ask, marked Q.
A few lines of code	A small cell that answers it, with a comment on each step.
The takeaway	One line on what the output shows.

Four habits run through every day.

Habit	What it means
Excel side by side	Wherever Excel has a function or feature for the job, it is shown next to the Python, formula included.
Errors on purpose	Code that fails is run for real, and the error is read line by line.
Exercises you can check	Each exercise has a check cell that tells you whether your number is right.
The AI check	Each day ends with code written in the style of an AI assistant. It has one bug. You find it by checking the output against a number you worked out yourself.

That last habit is the point of the course. You will use AI assistants to write code. Code that runs is not the same as code that is right, and the way to tell the difference is to know what the answer should be.

The courses

Each course has its own folder, its own course map, and its own list of what to install. Weeks start again at 1 in every course.

Course	What it covers	Status
1	Python Foundations: variables to pandas, charts, exploratory data analysis, text, and a project, in four weeks	Ready
2	Bond Math: price, yield, accrued interest, duration, DV01, convexity, spread, and DTS, built by hand in your own tested module, checked against Excel, and finished with a risk report, in three weeks	Ready
3	Market Data and Time Series: an API key kept secret, FRED's JSON, market calendars, resampling, changes, volatility, z-scores, percentile ranks, slopes, butterflies, SQL, AI-written code reviewed with tests, and a daily market snapshot, in three weeks	Ready
4	Machine Learning: regression on the spread cross-section, logistic regression, thresholds, decision trees, walk-forward validation, clustering, PCA, and a three-model capstone, each model tested on rows it never saw, shown in confusion heatmaps, curves, and formatted tables, and described, never forecast, in three weeks	Ready
5	Advanced Machine Learning: bagging, random forests, impurity and permutation importance, AdaBoost, gradient boosting, XGBoost, LightGBM, stacking, honest tuning, feature engineering on financial ratios, rare events, SHAP values, and a four-model capstone, with Course 4's test rows kept closed, shown in confusion heatmaps, curves, importance bars, and formatted tables, in three weeks	Ready
6	Neural Networks: tensors and gradients by hand, one neuron as a logistic regression, dense networks in Keras trained one setting at a time, losses, optimizers, batch normalization, dropout, early stopping, class weights, sequence models on the Treasury curve described, never forecast, and a capstone against LightGBM, all on the CPU, shown in loss curves, confusion heatmaps, ROC and precision-recall curves, residual plots, and formatted tables, in three weeks	Ready
7	NLP, RAG, and LLMs: regular expressions and tokens, TF-IDF, a text classifier on 8-K sections, word vectors, sentence embeddings and semantic search, zero-shot classification, EDGAR, requests to large language models through OpenRouter with the key kept secret, covenant extraction from real indentures with every number checked against the text, retrieval with citations, and a covenant extraction tool as the capstone, descriptive extraction only, shown in confusion heatmaps by split, embedding scatters, annotated cosine heatmaps, token and cost bars, and formatted tables, in three weeks	Ready
8	Implementing Agentic AI: the first tool request by hand, the agent loop with a step limit, tools from your own bondmath module, structured answers with every figure traced to a tool result, market and document tools, MCP, the OpenAI Agents SDK, an orchestrator with four specialists, memory, evals with answer keys decided first, guardrails and a budget, and a multi-agent credit desk as the capstone, every run replayed from saved responses so no key is needed, shown in bars of tokens, requests, and cost, tool-call counts, pass and fail grids, and formatted tables, in three weeks	Ready
9	Deployment: remotes, branches, pull requests reviewed by their diff, CI on Linux with GitHub Actions, what a server installs, a FastAPI bond service deployed to Vercel, the Course 8 desk served as a snapshot with no key on the server, Streamlit and Gradio front ends, a Hugging Face Space, Docker, request logs and monitoring, model files and drift, secrets, limits, and a cost meter, and the desk's brief as a public app as the capstone, all on free accounts, every notebook offline with the cloud steps shown as masked transcripts, shown in bars of transcript durations, sizes, request counts, and status codes, run histories, and formatted tables, in three weeks	Ready
10	Implementing RAG: five questions asked closed book and with a prompt alone, a filing as pages and items, a recursive splitter, embeddings from three pinned models, exact search with NumPy and FAISS, BM25 and hybrid search, a cross-encoder reranker, a tuning grid read on development questions, metadata filters across four documents, prompts that must cite, recall and MRR on a labelled set, LLM judges measured against a strict grader, index refresh and corpus drift, a retrieval service and its cost, and a cited question-answering tool over public filings as the capstone, every answer replayed from saved runs so no key is needed, in three weeks	Ready
Bonus	Build Your Own Coding Agent: a coding agent from an empty file on OpenAI's Responses API, the switch from Chat Completions field by field, three tools fenced to one workspace folder with none that runs code, a loop with a round limit and a rollback, a refusal rule sent before any request, bond math the agent writes checked against Excel, a Gradio chat page with your own name on it, and a private ZeroGPU Space of your own, every request replayed so no key is needed, shown in bars of the tokens of each request, the cost per task, the tool requests, and the Space's build steps, and formatted tables, in one notebook	Ready

The Bonus course, Build Your Own Coding Agent, closes the series.

The data

Every issuer, ticker, CUSIP, price, position, trade, and counterparty in the portfolio files is invented. The courses share one invented portfolio. No real company and no real security appears in any holdings file.

File	What it holds
data/holdings.csv	200 bonds: price, accrued interest, yield, spread, duration, DTS, par held, and market value
data/benchmark.csv	821 bonds from 150 issuers, with index weights
data/issuers.csv	One row per issuer: ticker, name, sector, and rating
data/ratings.csv	The rating scale, with a score, a bucket, and investment grade or high yield
data/positions.csv	A lean version of the holdings, for joining to the other tables
data/holdings_messy.csv	The holdings with problems planted on purpose, for the data cleaning lessons
data/holdings_messy_answer_key.md	Every planted problem, listed
data/trade_blotter.csv	374 trades in September 2026, in bonds from the holdings, with invented counterparties
data/project_holdings.csv	The course project's extract: the same desk one month later, as of 2026-10-30, with new problems planted in it
data/project_cover_note.txt	The cover note that comes with the project extract, with its control totals
data/project_answer_key.md	Every problem planted in the project extract. Read it after you finish
data/bondmath_excel_reference.csv	306 bond cases priced by desktop Excel: coupon dates, day counts, price, accrued interest, yield, duration, DV01, and bumped prices, for Course 2's tests
data/bondmath_excel_reference_tvm.csv	153 week 1 cases from Excel: FV, PV, RATE, EFFECT, NOMINAL, PRICE, and YIELD, each with the exact formula
data/course2_treasury_curve.csv	10 tenors of the invented Treasury curve the holdings were built on
data/course2_capstone_holdings.csv	The Course 2 capstone's extract: 201 rows, the desk as of 2026-11-30, with bond-math problems planted in it
data/course2_capstone_cover_note.txt	The cover note that comes with the capstone extract, with its control totals
data/course2_capstone_answer_key.md	Every problem planted in the capstone extract. Read it after you finish
data/course3_synthetic_spreads.csv	2,940 daily investment grade and high yield spreads in bp, one per date of the Treasury file. Synthetic: invented for Course 3, not ICE, Moody's, or any index
data/course3_capstone_extract.csv	The Course 3 capstone's vendor extract: 69 rows, 2026-07-01 to 2026-10-02, four Treasury tenors and the two synthetic spreads, with data problems planted in it
data/course3_capstone_cover_note.txt	The cover note that comes with the Course 3 extract, with its control figures
data/course3_capstone_answer_key.md	Every problem planted in the Course 3 extract. Read it after you finish
data/course4_excel_reference.csv	294 values computed by desktop Excel for Course 4's tests: LINEST statistics, p-values, intervals, and VIFs on the benchmark, Breusch-Pagan by hand, ridge in closed form, Solver's logistic fits on the card file, COUNTIFS confusion counts, and a trapezoid ROC AUC, each with its formula

The data is built to behave like a real portfolio.

Rule	What it means
Price and yield always agree	Each yield is a synthetic Treasury yield plus a spread, and each price is calculated from that yield.
One convention throughout	Fixed-rate bullet bonds, semiannual coupons, 30/360 day count, modified duration.
Identical for every student	Seeded scripts generate every file: tools/make_holdings.py for the portfolio, tools/make_blotter.py for the trades, tools/make_project_data.py for the project extract, tools/make_course2_data.py for the Course 2 curve and capstone, and tools/make_course3_data.py for the Course 3 spreads, FRED-format samples, and capstone. tools/make_excel_reference.py, tools/make_course3_excel_reference.py, and tools/make_course4_excel_reference.py wrote the Excel values once, through desktop Excel on Windows.
Checked	python -m pytest confirms the identifiers are valid and that each bond's price, accrued interest, and duration match its yield.
No real CUSIPs	Every CUSIP uses an issuer number in the range reserved for internal use, with a valid check digit.

Five files are real, public data, and they are not part of the portfolio. Their values are unchanged.

File	What it holds	Source and terms
data/treasury_par_yields.csv	Daily Treasury par yield curve rates, 2015-01-02 to 2026-10-02, 14 tenors	U.S. Department of the Treasury, Daily Treasury Par Yield Curve Rates. Listed by Treasury as public data under a public-domain dedication.
data/credit_card_default.csv	30,000 credit card accounts in Taiwan in 2005, with whether each defaulted the following month	Yeh, I. (2009). Default of Credit Card Clients [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C55S3H. Licensed under CC BY 4.0. Changes: the extra header row was dropped and the file was saved as CSV.
data/fomc_statements.csv	The 46 FOMC post-meeting statements from 2021-01-27 to 2026-09-16	Board of Governors of the Federal Reserve System, federalreserve.gov. The Board states that information on its website is in the public domain unless otherwise indicated. Only white space was changed.
data/polish_bankruptcy_5year.csv	5,910 financial statements of Polish companies, 64 financial ratios (Attr1 to Attr64), and whether each company went bankrupt within a year: the 5th-year file of the dataset, anonymised	Tomczak, S. (2016). Polish Companies Bankruptcy [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C5F600. Licensed under CC BY 4.0. Changes: 5year.arff from the UCI zip converted to CSV, the missing marker ? written as an empty cell, class decoded to 0 or 1, and every other value written as the shortest text that reads back to the same number. Nothing recoded, corrected, sorted, or dropped.
data/taiwan_bankruptcy.csv	6,819 rows of Taiwanese companies' financial ratios, 1999 to 2009: 95 features and whether each company went bankrupt, anonymised	Taiwanese Bankruptcy Prediction [Dataset]. (2020). UCI Machine Learning Repository. https://doi.org/10.24432/C5004D. Licensed under CC BY 4.0. Changes: none to the contents; data.csv from the UCI zip renamed and kept byte for byte.

The first three were retrieved on 2026-10-03, the Polish file on 2026-10-04, and the Taiwanese file on 2026-10-05. The scripts tools/get_treasury_yields.py, tools/get_credit_card_default.py, tools/get_fomc_statements.py, tools/get_polish_bankruptcy.py, and tools/get_taiwan_bankruptcy.py download each one again from the same source.

Course 3 adds three files derived from the public Treasury file. They hold its values, unchanged or computed from them, and no data downloaded from FRED.

File	What it holds	Source and terms
data/course3_fred_format_dgs10.json	43 observations of the 10 Yr, 2026-08-03 to 2026-09-30, in the FRED API's response format, with . for 2026-09-07	Values copied from data/treasury_par_yields.csv; the response format is FRED's structure, not FRED data
data/course3_fred_format_dgs2.json	The same for the 2 Yr	As above
data/course3_excel_reference.csv	190 market dates of 2026: the 2, 5, 10, and 30 Yr, and Excel's results for changes, rolling volatility, z-scores, percentile ranks, 2s10s, and 2s5s10s, with each formula	Computed by desktop Excel from data/treasury_par_yields.csv

Nothing pulled from the FRED API is stored in this repo. Course 3's live cells pull public Treasury series with the student's own key, and pull ICE BofA and Moody's series only to the student's own screen: those are licensed, held in memory, and never cached, written to a file, or committed.

Course 7 adds text: public statements, public filings written by real issuers, a public-domain word vector file, and outputs the course computed itself. Real issuers appear only as the authors of the documents, which are quoted for descriptive extraction only: what the document says, never a view on the issuer or its securities. The 46-row data/fomc_statements.csv is unchanged.

File	What it holds	Source and terms
data/course7_fomc_statements.csv	226 FOMC post-meeting statements, 2000-02-02 to 2026-09-16, 13 of them unscheduled	Board of Governors of the Federal Reserve System, federalreserve.gov, public domain unless otherwise indicated. Only white space was changed.
data/course7_indentures/	Three indentures as filed, each with its text: Boyd Gaming Corporation, 4.750% Senior Notes due 2031 (Exhibit 4.1 to the 8-K filed 2021-06-08); Levi Strauss & Co., 8-7/8% Senior Notes due 2016 (filed 2006-03-17); Lamar Media Corp., 6-5/8% Senior Subordinated Notes due 2015, Series C (filed 2007-10-16)	Public records on sec.gov, which states that information on its site may be copied and distributed without the SEC's permission. Kept byte for byte; the .txt files are the course's text of each.
data/course7_boyd_2021_indenture.pdf	The Boyd indenture printed to PDF by the course from the SEC copy, 132 pages	The SEC copy's text; the page layout is the browser's
data/course7_boyd_10k_2025.htm	Boyd Gaming Corporation's Form 10-K for fiscal 2025, as filed	A public record on sec.gov
data/course7_risk_factors.csv	99 Item 1A risk factors from the fiscal 2025 10-Ks of Boyd Gaming Corporation, Levi Strauss & Co., and Lamar Media Corp.	Public records on sec.gov. Only white space was changed.
data/course7_8k_items.csv	5,126 item sections of Form 8-K current reports filed 2019 to 2025, eight items, each labelled by the item its filer chose, the bodies cut at 300 words	Public records on sec.gov; a sample drawn by tools/get_8k_items.py with a fixed seed
data/course7_glove_6b_100d.npz	GloVe 6B word vectors, 44,558 words at 100 numbers each	Pennington, Socher, and Manning (2014), nlp.stanford.edu/projects/glove, Public Domain Dedication and License v1.0. A subset of the published file.
data/course7_8k_minilm.npy, data/course7_8k_zero_shot.csv	The two pinned models' outputs on the 8-K sections, computed once on one thread	The course's own, under the repo's MIT license
data/course7_llm_responses/	66 responses of large language models to the course's week 3 requests, one file per request, recorded on 2026-10-08	The course's own, under the repo's MIT license. No key and no request header is stored.

The scripts tools/get_fomc_history.py, tools/get_indentures.py, tools/get_edgar_filings.py, tools/get_8k_items.py, tools/make_indenture_pdf.py, tools/make_glove_subset.py, and tools/make_8k_model_outputs.py made these files on 2026-10-07; the EDGAR scripts declare a contact in their requests and stay under the SEC's limit of 10 requests a second.

Course 8 adds no external data. Its agents read the files above (the synthetic holdings, the Treasury curve, the synthetic spreads, and Course 7's statements, indentures, and 8-K sections), and it adds three files of its own.

File	What it holds	Source and terms
data/course8_eval_set.csv	30 desk questions with answer keys decided before any request: 20 answerable, 3 whose answer is not in the data, 5 adversarial, and 2 with an instruction planted in a tool result	The course's own, under the repo's MIT license. The keys are computed from the committed files.
data/course8_synthetic_desk_memo.txt	A 13-line desk memo about the synthetic book, with one planted instruction for the lesson on prompt injection	Synthetic: written for the course, under the repo's MIT license. It names no real issuer.
data/course8_agent_runs/	350 requests and responses of the course's agent runs, one file per request, recorded on gpt-4.1-mini from 2026-10-08 to 2026-10-11	The course's own, under the repo's MIT license. No key, no request header, and no account identifier is stored, and every tool result inside a request is computed from the committed files.

Course 9 adds no external data. Its notebooks read the files above (the synthetic holdings, the Treasury curve, Course 2's Excel reference values, and Course 8's eval set and the capstone's saved runs), and it adds one synthetic file and one folder of its own.

File	What it holds	Source and terms
data/course9_request_log.csv	19,664 requests to a bond and desk service over 30 days, 2026-09-03 to 2026-10-02 UTC: path, status, latency in ms, bytes out, and whether the request was refused, never a body, query, header, or client address. Three features are planted for the lessons: a slow day on /dv01, a burst of 422s from one malformed client, and a refused share on /ask	Synthetic: invented by tools/make_course9_request_log.py with a fixed seed; no real service, client, or request produced it. Under the repo's MIT license.
data/course9_transcripts/	75 masked transcripts in 11 day folders, recorded on 2026-10-12: what GitHub, Vercel, Hugging Face, and Docker printed when the course ran each cloud step on its own accounts	The course's own, under the repo's MIT license. Account names, emails, addresses, and ids are masked to student and <id>; no token is stored.

Course 10 reuses Course 7's Boyd 10-K and three indentures (the files above) and adds one PDF print, one new public filing, and four folders and files of its own.

File	What it holds	Source and terms
data/course10_boyd_10k_2025.pdf	Boyd Gaming's Form 10-K for fiscal 2025 printed to PDF by the course from the SEC copy: 113 pages; the text is the SEC copy's, the page layout is the course's print	Real, a public record on sec.gov, quoted for descriptive extraction only; the print is the course's own
data/course10_boyd_10q_2026q2.htm	Boyd Gaming's Form 10-Q for the quarter ended June 30, 2026, filed 2026-07-30, as filed (inline XBRL)	Real, a public record on sec.gov ("Information presented on sec.gov is considered public information and may be copied or further distributed"), quoted for descriptive extraction only
data/course10_questions.csv	75 questions with answer keys written first from the documents' pages, each key a quote with its page, item, or section: 5 for day 1, 25 development, 20 test, 25 capstone, including 8 the corpus cannot answer and 3 requests for a view	The course's own, under the repo's MIT license
data/course10_embeddings/	8 sets of embeddings (16 files, 14.4 MB) of the chunk tables by MiniLM, bge-small, and gte-small at pinned revisions, each with a sidecar naming the model, the chunker, and the hash of the chunk texts	The course's own derived data, computed once on one thread, under the repo's MIT license
data/course10_rag_runs/	352 requests and responses of the course's answers and judge verdicts in 9 day folders, one file per request, recorded on gpt-4.1-mini on 2026-10-12 and 2026-10-13	The course's own, under the repo's MIT license. No key, no request header, and no account identifier is stored.
data/course10_human_labels.csv	A reviewer of the course's groundedness and relevance labels on day 12's 45 answers, made from the question, the answer, and the chunks before any judge's verdict was shown	The course's own, under the repo's MIT license

The Bonus course adds no external data and reads no file above. Its agent works in a fictional workspace that its own module writes at run time: README.txt, a riddle, and a bonds.csv of six bonds from six invented issuers, with no CUSIP, labelled fictional in the folder and never a file in data/. It adds two folders of its own.

File	What it holds	Source and terms
data/bonus_agent_runs/	29 requests and responses of the course's agent on OpenAI's Responses API (gpt-6.1-sol), one file per request, recorded on 2026-10-12	The course's own, under the repo's MIT license. No key, no request header, no account identifier, and no absolute path is stored, and every tool result inside a request comes from the fictional workspace.
data/bonus_transcripts/	7 masked transcripts, recorded on 2026-10-12: what Hugging Face printed when the course created its private Space, built it, sent it three requests, ran it once without the ZeroGPU placeholder, and deleted it	The course's own, under the repo's MIT license. Account names, addresses, and ids are masked to student and <id>; no token is stored.

This product uses the FRED® API but is not endorsed or certified by the Federal Reserve Bank of St. Louis.

What is in this repo

python-fixed-income/
  course1_python_foundations/
    README.md           the course map
    requirements.txt    what the course's notebooks need
    week1/ ... week4/
      day1/ ... day5/   the notebook, its solutions, and its overview PDF
  course2_bond_math/
    README.md           the course map
    requirements.txt    what the course's notebooks need
    week1/ ... week3/
      day1/ ... day5/   the notebook, its solutions, and its overview PDF
  course3_market_data/
    README.md           the course map
    requirements.txt    what the course's notebooks need
    week1/ ... week3/
      day1/ ... day5/   the notebook, its solutions, and its overview PDF
  course4_machine_learning/
    README.md           the course map
    requirements.txt    what the course's notebooks need
    week1/ ... week3/
      day1/ ... day5/   the notebook, its solutions, and its overview PDF
  course5_advanced_ml/
    README.md           the course map
    requirements.txt    what the course's notebooks need
    week1/ ... week3/
      day1/ ... day5/   the notebook, its solutions, and its overview PDF
  course6_neural_networks/
    README.md           the course map
    requirements.txt    what the course's notebooks need
    week1/ ... week3/
      day1/ ... day5/   the notebook, its solutions, and its overview PDF
  course7_nlp_and_llms/
    README.md           the course map
    requirements.txt    what the course's notebooks need
    week1/ ... week3/
      day1/ ... day5/   the notebook, its solutions, and its overview PDF
  course8_agentic_ai/
    README.md           the course map
    requirements.txt    what the course's notebooks need
    week1/ ... week3/
      day1/ ... day5/   the notebook, its solutions, and its overview PDF
  course9_deployment/
    README.md           the course map
    requirements.txt    what the course's notebooks need
    week1/ ... week3/
      day1/ ... day5/   the notebook, its solutions, and its overview PDF
  course10_rag/
    README.md           the course map
    requirements.txt    what the course's notebooks need
    week1/ ... week3/
      day1/ ... day5/   the notebook, its solutions, and its overview PDF
  course_bonus_coding_agent/
    README.md           the course map
    requirements.txt    what the course's notebook needs (Course 9's file)
    week1/day1/         the notebook, its solutions, and its overview PDF
    deploy/             the three files a Hugging Face Space receives
  data/                 the invented portfolio and the public datasets, shared by every course
  tools/                the scripts that build or download the data, and the notebook banner
  tests/                checks on the data
  assets/               fonts and images

License

Code is released under the MIT License. The notebooks' written content, the slide overviews, and the videos are released under CC BY 4.0: you may share and adapt them, including commercially, as long as you give credit.

The public data files and documents keep their own terms, listed in The data. They are not relicensed by this repo.

The fonts in assets/fonts are under the SIL Open Font License, and their license files sit beside them.

---

Created by Jeff Lenamon