Guest Post: A More Granular Look at 2013 (weekday) OTP, Part I

Every month, we receive aggregate OTP data from the LIRR, by branch and by time of day. Sometimes, this data seems unbelievable, other times it's a bit disturbing. Sometimes we just need to prove to our bosses that the train was in fact late. New York City Transit has a verification letter system, LIRR and Metro-North have an even simpler system, but it makes you wait a week. Launched in mid-2010 under Jay Walder's push for MTA data transparency, the delay info site lists every single delayed or canceled train on every single day, special schedules and track work programs included.

Whenever one looks at data, the larger units will always tend to obscure the details. Say a month had a particularly terrible performance. Was that the result of a couple terribly poor days, or was it a little here, a little there? If it was rush hours, was it the peak of the peak or a really early train with comparatively lighter ridership? What about the individual train that you ride every day? Given that this granular data exposes this all in gory detail, I decided to do some analysis.

Notes: 

• A late train is defined by national industry standard as a train that arrives at its final terminal more than 5:59 after it's scheduled arrival time. There's no way around that.
• To avoid complications related to trackwork, I only looked at weekdays. This also avoids the most significant weather event of the year, Winter Storm Nemo.
• I tried to include all other special schedules I could think of (ESA, etc.), but a couple trains may be off here and there. Notably, Belmont Park service is excluded.
• Because of this, the data will not match up exactly with MTA official data, nor with actual operations.
• Trains may make up or lose time en route. An eastbound Montauk train incurring a 20min delay at Babylon is counted no differently than one incurring a 20min delay at Amagansett, despite impacting far more passengers.

Measuring On-Time Percentage

Before I begin the analysis, it's worth giving a mention as to the different ways we can measure OTP, which stands for on-time performance.

• Percent of trains on time: This is the most common measurement. It works well when we want to see the condition of a system over a large period of time, with lots of trains to look at. The downside is that it treats a train that's 6 minutes late the same way as one that's 40 minutes late, even though to the passenger they're two totally different things. It's also not really an accurate measurement when looking at very small sets of data, since one late train will appear to create an enormous impact. Cancellations and terminations count as late trains. Of note, the LIRR's goal on all branches is 95.1%. 
• Delay minutes: This measures the total number of minutes that trains were delayed. By definition, it excludes cancellations and terminations, so that aspect of customer impact is not reflected. Also, a single trespasser or vehicle strike could result in 100 or more minutes of delay to one particular train, creating an unusually high value. Which leads us to...
• Average delay: Just like delay minutes, this excludes cancellations and terminations. It's helpful for figuring out trains that are late often but not by much (which is usually a sign that schedule adjustments are needed!), or how badly most people were impacted. Just like any average figure, the number of incidents in the average must be considered.
• Passenger-minutes: This is, in my opinion, the "gold standard" for OTP measurement, as it measures actual customer impact. But it's a very difficult proposition on the LIRR, with such close rush hour headways, passengers getting on the first train they see, passenger extras, varying loads east and west of Jamaica, etc. Later on I will make an attempt at factoring in load, but the complexities of true passenger-minutes won't be explored here.

On to the data. You can find the full table at the end of the post. I will go from largest unit to smallest unit.

Days

The 5 worst days of 2013, by on-time percentage, are shown in this table:

Date	Cancel- lations	Term- inations	Late trains	Delay minutes	Scheduled runs	OTP	What happened?
Jan 31	18	16	164	2019	729	72.8%	High winds, leaning power pole
Mar 19	23	11	137	1925	728	76.5%	Derailment cleanup at Forest Hills, trespasser strike in Mineola
Jun 18	45	17	101	1074	737	77.9%	Derailment cleanup in Penn Station
Sep 5	15	9	136	1847	739	78.3%	East Side Access work gone wrong
Aug 19	13	6	135	3087	738	79.1%	Third rail problems in an East River Tunnel

Note: These are not the worst 5 by delay minutes, although 8/19 and 1/31 are 2nd and 4th, respectively.

There are some things in common. All of them involved incidents on the Main Line (the leaning power pole on 1/31 was just west of Hicksville), where one incident can have a drastic impact on many, many trains in both directions and flexibility is extremely limited. Four involved disruptions west of Jamaica, where the impact is even more pronounced. Only one of these days had inclement weather as a primary factor, two had Amtrak at fault, and two had causes that were, in theory, preventable by the LIRR. Some had multiple factors, and all had OTP's below 80%, which, given the circumstances, isn't that terrible. Both derailments of 2013 are on the list.

For comparison, here's the 6 best days (days 3-6 are a tie, and no cancellations occurred on any of these days):

Date	Terminations	Late Trains	Delay minutes	Scheduled runs	Percent on time
Nov 5	0	5	49	738	99.3%
Oct 14	0	9	71	737	98.8%
Sep 26	0	11	121	739	98.5%
Aug 13	0	11	98	734	98.5%
May 7	1	10	105	728	98.5%
Feb 4	0	11	102	728	98.5%

Note: These are 1st, 2nd, 12th, 4th, 6th, and 5th best by delay minutes, respectively.

There were no perfect weekdays in 2013. The first two days are Election Day and Columbus Day, respectively, potentially lowering ridership slightly. With the exception of 9/26, the rest are Mondays or Tuesdays. All were precipitation-free, otherwise, I don't see a trend in these days, they just seemed to be when the stars crossed for the LIRR and its riders.

Rush Hours

Because the rush hour involves such heavy loads, ridership, and congestion, I'm going to list the worst 5 in both measures. Throughout this section, rush hour includes only those trains classified by the LIRR as peak trains, not reverse peak runs.

Date Rush	Rank %/Min	Cancel-lations	Term- inations	Late Trains	Delay  minutes	Scheduled  runs	OTP	What  happened?
8/15 AM	1/1	4	2	84	2952	141	36.2%	Disabled train blocking JAY
5/16 PM	2/6	22	0	56	1316	130	40.0%	Power loss in HAROLD
8/5 AM	3/4	3	5	76	1358	141	40.4%	East River Tunnel work ran long
9/5 PM	4/14	13	1	63	795	130	40.8%	East Side Access work gone wrong
8/19 AM	5/2	10	6	68	2153	144	41.7%	Third rail damage in East River Tunnel
4/10 AM	15/3	5	0	55	1596	144	58.3%	Disabled train west of Hicksville
11/20 AM	16/5	0	1	59	1321	144	58.3%	Disabled train at Carle Place

There are lots of things to be gleaned here. Let's start with the basics: there are 2 PM rush hours and 5 AM rush hours on this list. Though in general, OTP for the PM rush is much lower than that of the AM rush, it appears that when things go down the toilet the AM rush makes it further. The reasons are clear: there are more trains in the AM than the PM, and ridership is significantly higher. There are more flyers, expresses, and hours of single-directional operation on the busiest lines, significantly limiting operating flexibility.

The next thing that I notice is the number of cancellations. The LIRR is eager to cancel eastbound trains out of Penn when things go wrong. But when something fails out west during the morning rush, the railroad seems very reluctant to cancel (or divert) trains originating from points east, instead opting to run them through with significant delays. The worst rush hour of the year (on the worst day of the year by minutes) is a good example: on August 15, train 1001, the 5:59am Massapequa Park departure, broke down and blocked 3 of the 4 tracks west of Jamaica that lead to Penn Station. Rather than immediately start canceling trains originating at eastern terminals within 20-30 minutes of the incident, the railroad decided to let them go and started stacking them as far as Hicksville and Islip, with the vast majority of the 144 scheduled AM peak trains sitting one behind another waiting to snake through the one track westward. In fact, disregarding Port Washington Branch service, OTP for August 15's AM rush would be 30.6%. Counting only NYP/HPA/LIC trains that pass through Jamaica, it would be yet lower.

Delays are not just a direct ratio of the ratio of trains to available capacity, as the ratio grows, delays grow in a non-linear fashion. November 20th and April 10th are also good examples of this, as well as the significant delays a few weeks ago, when a Ronkonkoma train had equipment problems at Bellerose in the middle of an ice storm. The lack of the Main Line Third Track is directly responsible for allowing these delays, some of which reached 40 to 60 minutes, to build. When you have trains flying at 80mph every 3 minutes neck to neck one after another down two tracks, attempting to send them over a switch at reduced speed down one track with different stopping patterns will lead to a massive traffic jam. When things are fixed and the tracks reopen, the resulting gridlock takes a long time to clear, even longer if the clock has struck 8:10am and a line of eastbounds start to form. South Shore and Port Washington trains aren't impacted too badly, so OTP isn't in the 30's, but the delay minutes rack up fast.

Another major disruption-causer that stands out is Amtrak. Of these 7 disastrous rush hours, 3 (43%) were partly or completely Amtrak's fault. However, for all of 2013, Amtrak caused only slightly more than 12% of delay incidents. This disparity can be explained in a few ways:

• Amtrak controls the biggest station the LIRR serves, New York Penn, so any issue there ripples across the system.
• There are 4 East River Tunnels, so when something goes wrong during the off peak, it's not a big deal, but once rush hour hits, the delays mount rapidly.
• Repairs in a tunnel take a lot longer than repairs on the surface.
• HAROLD/F and JAY/HALL, some of the busiest interlockings in the country, congest easily, so delays build and spread much more than if the delay occurred elsewhere on the system.

Lets take a break from the doom and gloom of peak operations recoveries and look at the 100% rush hours of 2013, of which there were 36:

• AM: 1/8, 1/15, 2/20, 3/5, 4/9, 4/16, 4/23, 4/24, 4/25, 5/1, 5/3, 5/6, 5/9, 5/14, 5/29, 6/4, 6/6, 6/12, 6/27, 6/28, 7/5, 7/18, 7/26, 8/13, 8/20, 9/26, 10/1, 10/14, 11/5, 11/6, 11/21, 12/4, 12/9
• PM: 3/5, 4/8, 4/23

Note that March 5 and April 23 show up in both lists! Most are on Tuesdays and Wednesdays for some reason. The low number of perfect PM rushes has a variety of explanations. Systemwide, eastward connections at Jamaica need to be held when AM westbounds don't need to be, and unloading large numbers of passengers can take longer than loading them. From a Main Line perspective, the Babylon Branch retains its reverse peak track for the entire rush while the Main Line only loses it for about 45 minutes, less than half the time of the AM peak. That means more trains must be sent down one track, with any extended dwell time on a local stacking expresses up behind it.  On the Babylon Branch, one small hiccup can cause the rest of the trains for the rush hour to be late. Plus, reverse peak PM is busier, with people seeking to go into the city for a night out, in addition to those heading home from work.

That's the entire rush hour. But what about within the rush hour? These graphs show trains arriving at Penn and their percentage OTP, relative to their scheduled arrival time. First, the morning:

Note the significant dip around from 8:45-9:15am. There are several hypotheses I can think of for the graph looks the way it does, and the real answer is probably some combination of them. The AM rush starts on what is basically a blank slate. There is no day of service behind it that could cause lingering issues going into the mess. So it starts out performing nicely. The first hour or so of the service is lightly used, so dwell time and slow loading don't show up much, leading to good OTP. The next hour, ridership rises, and the Mainline becomes unidirectional. Congestion begin to manifest itself as minor, negligible delays. As we enter the 8am hour, ridership continues to rise and peaks at 8:30am. But the Main Line starts accepting eastbounds, and the level of service drops. Loads rise, and dwell times increase. More importantly, these delays are on top of the little things that are already there. It gets worse and worse, until ridership starts to lighten up shortly after 9am. From that point on, service falls precipitously, and a recovery period begins. The first westbound off peaks, though not shown, don't fare too nicely either, as they are often dependent on those first reverse peak deadheads for their equipment.

Now the PM:

Like above, the peak of the delays tends to lag behind peak passenger load in the 5:30 hour. There is a greater range of values here, because this is endpoint OTP that is recorded at many different terminals, and different lines have different propensities for delay. The initial OTP is lower, since there are more trains rolling when the clock strikes 4.

When I made these graphs, I expected Port Washington Branch service to trend similarly to other branches, but be higher the rest because it doesn't pass through congested Jamaica. It turns out that's not true, perhaps because HAROLD is a greater chokepoint in the morning, and Great Neck trains are often the first to go when something screws up in the East River Tunnels or thereabouts in the evening.

Despite some of the ugly numbers, they should look to good to you. Why? Because in January 2014, each one of those lows for the year was beat. Mother Nature doesn't care about on time performance in the least, in fact you can say she worked her hardest to screw it up. More in a future post.

In a post later this week, I will take a look at each individual train and some of the worst offenders, including what those outliers are.

Included below is the full table of OTP data, calculated by day and rush hour.  You can switch between different sorting styles by using the tabs at the top of the box.  To view the spreadsheet in a different window, you can click here.

(Note: If you subscribe to the site by e-mail or RSS, you may have to view this post in your web browser to see the embedded table)

A special thanks to Sunny Zheng for all his work on this post!